Skip to main content

Readiness for a stable 1.0

The target is a complete, extensible coding agent that can be used to develop Compositor itself and other projects. A release number, a large tool catalog or a passing unit suite does not establish that outcome. This page tracks the required evidence; it is not a declaration that 1.0 is ready.

Current development checkpoint — 2026-11-15​

On 2026-11-16, npm audit --omit=dev at 447e6fd reported five low-severity dependency findings, all in one chain from the documentation workspace through Docusaurus and SVGR to deepmerge 4.3.1. The upstream advisory lists no patched version. The installed SVGR plugin's distributed entry does not import deepmerge; this observation is limited to that entry and is not a general proof that no other build path can load it. The production audit covers the documentation workspace as well as the application; it is not a scan of the contents or execution paths of each native archive. Do not treat npm's suggested Docusaurus downgrade as a verified fix. The finding remains open.

GitHub reports that the earlier build-only run 43759318216 for c2fd409 completed verification and all four native build jobs successfully; publication was skipped. This predates the /btw recovery gate and the later fixes. The reported job results do not certify those later changes or establish 1.0 readiness.

Build-only run 43762410510 for 938edb9 failed the installed Linux ARM64 /btw scenario: its side draft was saved as BTW_SIDE_DRAF although the input trace contains the final T. The retained failure report and terminal trace are in dist/candidate-938-arm-failure/opentui-btw-linux-arm64/. Later changes include retry handling for an earlier locally typed field value and input-admission ordering. A controlled production-renderer regression delays the final edit, publishes its previous prefix and switches focus; with the 938edb9 native client it retains DRAF, while the current client retries and retains DRAFT with Main held running. The test observes rejection of the old revision and acceptance of the retry. That hybrid source experiment establishes one possible mechanism, not the exact event ordering of the ARM64 failure. The candidate completed with the other three platforms passing and publication skipped. Current-source four-platform certification remains required; build-only run 43766867589 for c119f4e has started verification. The hybrid and current test outputs are retained under dist/draft-revision-comparison-20261116/.

An installed Linux x64 OpenTUI /btw probe reproduced a stale side Reply field: the draft reached private storage after switching to a held Main, but reopening the side window showed the prior empty value. Agent-hub now republishes its current draft after a valid input change. The same controlled flow passes on source and a locally rebuilt binary, including window closure and restart; reports are dist/visual-opentui-btw-native-audit-c2fd409/ and dist/visual-opentui-btw-native-agent-fix/. This is one loopback-provider scenario, not proof for physical terminals or other platforms.

An isolated browser-backed PTY check on a locally installed Linux x64 binary from c2fd409 created two managed agents, retained separate drafts across restart, and exercised delayed and conflicting field updates. Its loopback provider received one explicit reply; the report and executable hash are in dist/visual-opentui-agent-drafts-native-c2fd409/. This is not a /btw command test, a physical-terminal check, or evidence for other platforms.

An extracted btw 2.0.0 package on the source Ink PTY host passed concurrent conversation and cancellation checks but once timed out after restart while waiting for its new reply. The input trace shows PageDown immediately after submission; the failure screen retained earlier side history. An unchanged repeat passed. The scenario now gates the fixture response and observes the side agent running and then idle before scrolling. Two local gated runs passed; these observations do not establish the cause of the original timeout or verify a native OpenTUI archive. Reports are under dist/visual-btw-package-c2fd409/ and dist/visual-btw-gated-c2fd409/.

An installed Linux x64 auto-review diagnostic used Qwen3.5 2B Q4_K_M through a local direct CPU executable and a loopback transport. With grammar-constrained JSON, the real bundled plugin allowed an explicitly requested npm test and denied the same command when the user had prohibited all commands. Both cases preserved an unrelated draft; the denied command did not run. The allowed case's model explanation incorrectly described the command as non-destructive, while the fixture's test script wrote a marker. This is two known cases, not evidence that the model reliably infers arbitrary script effects or is safe for unattended review. The diagnostic used no paid calls and no OS network sandbox; its script, responses, executable/model pins and report are in dist/local-auto-ci-5af22af/. An isolated copy of that installed plugin was then given an explicit instruction not to infer script effects from npm test. On the same two cases, the model again allowed the requested command and denied the prohibited one, but still asserted that the requested command was non-destructive without inspecting its script. The modified plugin SHA-256 was 669b64b0a618490f1b6ce2a0730cff20969c9354b48942b764bd8e82c93d057a; its local report is dist/local-auto-ci-variant-24a0a80/report.json. This single-model prompt experiment did not establish a repair and was not shipped. A further isolated variant added the exact synthetic test script and its fixture file contents to the review payload. The loopback transport verified that the evidence reached the model. The same model still called the requested run non-destructive despite the supplied file-write operation; it again denied the explicitly prohibited run. The modified plugin SHA-256 was dca18ceff7de06b88ca5a99098feb2809d0ccb61f22e64d8429ea530dded3e2a and the verified report is dist/local-auto-ci-evidence-d579355/report-verified.json. The evidence was synthetic: this experiment neither read user project scripts nor established a safe mechanism for disclosing or keeping them current.

In a local serverless onboarding experiment at af30f14, a small model received a synthetic, authoritative plugin-state snapshot but incorrectly said disabled test reports prevented test execution. It did not claim to perform the requested activation, and no action tool was available. The three-turn report and limits are recorded in the CPU diagnostic. Verified state in the prompt alone does not establish reliable recommendations.

A local Qwen3 1.7B CPU run against 89ec1a4 found and edited the known multi-file case, but reached its 12-request limit with invalid JavaScript and no successful verification. The independent grader failed; protected files survived and the owned server exited. The report is dist/local-cpu-navigation-89ec1a4/report.json. This is one known task, not a general model-quality measurement.

Node 26 CI for 89ec1a4 failed a restored custom-editor scenario. Its final screen shows the submitted text retained as an unaccepted draft. An existing held-startup test confirms that input is rejected while a replacement starts; the CI trace does not establish whether that exact condition caused the failure. Main now exposes Restoring during the switch, and the scenario waits for the actual ready state before submitting. The original CI run remains failed.

The ignored-parent discovery fix introduced a separate exclusion regression: re-rooting Git could expose files covered by path-specific enclosing ignore rules. An installed Linux x64 probe reproduced this in both glob and semantic search. The revised implementation retains specific ancestor .gitignore, repository and global exclusions, including when the selected root is reached through a symlink; conservative filtering can hide additional untracked files. The same installed probe passes locally after rebuilding; reports are dist/ignored-all-excludes-before-fix-20261116.json and dist/ignored-all-excludes-native-final-20261116.json. Other platforms and public release artifacts have not been verified for this change.

An unknown model tool call previously disappeared from provider context after session resume. Journal-only call metadata now preserves the original arguments and rejection, including for calls omitted by a batch strategy. Source regressions cover mixed and parallel batches without re-running tools or altering hook input. An installed Linux x64 process-death fixture fails on the earlier binary and passes on a locally rebuilt binary; reports are dist/rejected-call-before-fix-20261116.json and dist/rejected-call-native-fixed-20261116.json. Older journals lack the metadata; other platforms and external providers remain unverified for this change.

An isolated Linux x64 release smoke first failed discovery when its temporary project was itself inside the checkout's ignored dist directory; the same smoke outside that repository passed. A focused installed-binary regression then reproduced the failure in a fresh parent Git repository whose ignore rule covers the selected project. The revised source retains tracked files and uses a separate temporary Git index to enumerate the explicit root under its own and repository/global ignore rules. Local tests also cover a colliding parent index name and agent ignores. The prior binary fails and a locally rebuilt binary passes the same installed discovery probe; reports are dist/ignored-parent-native-before-fix-20261116.json and dist/ignored-parent-native-20261116.json. This does not establish behavior on other platforms or a public release. A follow-up installed semantic query in the same ignored-parent fixture finds visible_probe while excluding an .agentignore symbol; the prior binary fails both glob and semantic checks, as recorded in dist/ignored-parent-native-before-semantic-20261116.json.

A locally built Linux x64 installed CLI was killed after a controlled plugin appended a file and before its result was recorded. On resume, its second loopback model request retained the original task and an explicit unknown-outcome result; the effect was not repeated and an unrelated draft survived. The report is dist/effect-crash-native-20261116.json; this is a single controlled effect, not proof for arbitrary services, other platforms or multi-hour sessions.

The effect-boundary regression starts a separate core process, waits until a controlled tool has modified a file and its call is journaled, then sends SIGKILL before any result is recorded. A resumed core retains the original task, reports the effect outcome as unknown to the next scripted provider request, and does not replay the available tool. The test uses a local append-only journal and fixture effect; it does not prove recovery for arbitrary external services or multi-hour work.

Build-only run 43750368231 for ef87cbb failed the Linux ARM64 installed delegated-request scenario: its exact model-call ledger recorded three detached requests where the scenario expected four. The retained trace shows the final detached launcher and its cancellation, but no fourth provider request. The test previously closed the application immediately after launching that child while requiring its model request in the final ledger. It now requires the new provider request and a visible child question before closing, so the shutdown assertion exercises an actually pending child. This addresses a possible timing race; it does not establish the original ordering or turn that failed archive run into a pass. Evidence is retained under dist/candidate-ef87cbb-linux-arm-failure/.

An isolated local Linux x64 archive built from clean b6ad606 passed the SIGKILL queued-input and interrupted-retry checks after extraction into a separate directory. The installed executable SHA-256 is 9064329369267e3a66fb587975a02061e498b06d998b3e7f2dc37fa0af774393; reports and screenshots are under dist/native-recovery-sigkill-20261116/ and dist/native-recovery-interrupt-20261116/. This checks one locally built platform with a loopback provider, not the public release or the other platforms.

A local browser-backed PTY check now kills the actual OpenTUI CLI with SIGKILL only after a queued message or steering notification is saved in its session journal. The bridge observed exit status 137 in both runs. After a fresh start, each input appeared for explicit review; no new model request occurred until /recover retry 1. Reports are in dist/recovery-sigkill-20261116/ and dist/steering-sigkill-20261116/. These source checks use a scripted provider and do not establish multi-hour behavior, native-terminal compatibility or recovery at arbitrary effect boundaries.

A separate mixed-input probe killed the held process after both a steering intent and a queued intent were present in its journal. The resumed session listed them in that order without a new model request; explicitly retrying the first sent exactly one request and left the second available. The source and previously built local Linux x64 archive passed with a scripted loopback provider; reports are in dist/visual-mixed-sigkill-retry-source-fixed/ and dist/visual-mixed-sigkill-retry-native-fixed/. This tests one controlled crash point and does not establish durability for arbitrary power loss, journal backends, other platforms or multi-hour sessions. CI and the native release matrix now include the scenario, but their results must be checked separately.

Build-only run 43749137231 for 8c789e6 failed its Linux x64 installed OpenTUI plugin placement check. The retained form shows Network proxy selected and moved to workspace 2 while the test expected Todos. The input trace contains two rapid right-arrow presses; the exact frames between them were not retained. The scenario previously accepted any terminal frame as confirmation of a selection key, which could allow another key before the selected field changed. It now waits for that field's value to change and checks the chosen plugin immediately before Apply. The revised source PTY scenario passes; this addresses a possible timing race, but does not establish the cause or make the failed installed-archive run pass. A new candidate remains necessary. Failure evidence is under dist/candidate-8c789e6-linux-failure/.

A local CPU-only Qwen3 1.7B run started from clean 8c789e6 did not solve the known multi-file-bug case within 12 model requests. It searched for task phrases, repeatedly attempted a nonexistent file despite navigation hints, then listed the project and read the relevant source before reaching the limit. No successful project verification was recorded, and the independent behavior check failed; the protected package and check files were preserved. The owned loopback server exited, and the retained report is dist/local-cpu-multifile-8c789e6/. This single known-case diagnostic used no paid requests and does not establish general model quality. The worktree was edited after the process started, so the run is not exact-revision certification.

The workflow ownership regression now uses a separate OS process to hold an actual workflow lease. An observer cannot pause or stop its active run or overwrite its completed journal from a stale snapshot. A second case kills the owner, observes a paused recovery without an automatic retry, and resumes explicitly. Both use scripted workers and local temporary files; they do not establish multi-hour recovery, real-model quality, or cross-machine coordination.

Build-only run 43724725266 for 69078aa passed source verification and all four Linux/macOS x64/ARM64 jobs, including the previously failing Intel Main-copy scenario; publication was skipped. The downloaded quality and four installed-runtime evidence archives match GitHub's recorded SHA-256 digests. The quality archive has 2702 passing tests, five skips, and 32 certificates with 405 passing results and 1347 matching evidence references. Each platform archive has 13 successful OpenTUI reports (88 checks) and nine successful smoke reports. These retained reports do not independently verify all four binary artifacts or cover later auto-review, form and workflow changes, physical terminals, or general 1.0 readiness.

The generated plugin-release PR's Node 26 CI run 43726191303 failed the configured OpenTUI application scenario while waiting for the next model response. Its retained input trace contains the typed prompt and Enter; the final screen still shows that prompt as a draft. The exact running state at Enter was not recorded. Main intentionally keeps ordinary input as a draft while the preceding turn is running, and answer text can appear before that turn settles. The scenario now waits for Main's ready state before the next submit. An isolated loopback test that deliberately delays the preceding response's finish reproduced the same missing-response failure without the wait and passed with it; both PTY reports are retained under dist/opentui-app-no-ready-control-20261116/ and dist/opentui-app-ready-delayed-20261116/. This demonstrates a possible race, not the precise cause of the original failure. The failed run remains failed.

Build-only run 43722164397 for f72b017 passed source verification and the Linux x64, Linux ARM64 and macOS ARM64 jobs, but failed the macOS Intel Main-copy migration scenario while parsing the fork fixture's JSON. The exact bytes read at failure were not retained. The fixture previously wrote its active and fork JSON files in place while the observer could read them; it now publishes complete files by rename. Concurrent JSONL reads in this scenario also use the complete-line parser. These changes address possible partial reads without establishing the cause of the original failure. The failed candidate remains failed; a fresh four-platform run is required. The revised fixture subsequently passed all eight Main-copy checks in both native and selected-controller compositions against the downloaded Linux x64 binary, with a natural PTY exit. That local run does not establish Intel macOS behavior. Evidence is in dist/candidate-f72b017/copies-fixture-repair/. Its source CI passed on Node 22 and 26. The downloaded quality and three successful installed-runtime evidence archives match GitHub's SHA-256 digests; the 32 certificates contain 405 passing results and 1347 matching evidence references. The failed Intel evidence archive also matches its recorded digest. An independently downloaded Linux x64 binary artifact matched its digest, reported version 0.8.0, and passed nine local PTY navigation checks including keyboard retiling and restart. This does not qualify the other binaries or establish daily usability, physical-terminal coverage or 1.0 readiness.

Source CI for 321a8c8 failed its Node 26 history-resume test: saved focus differed from the baseline when the assertion ran. The captured log does not establish the exact ordering of session startup, focus restoration and the storage read. The test now gates replacement startup and waits for the host's actual automatic Main selection before checking saved placement, while retaining its original placement, closed-window and split-ratio assertions. This addresses a possible premature assertion; it does not turn the failed CI run into a pass or prove its cause.

Build-only run 43709559231 for 9802f1b reports successful verification and all four Linux/macOS x64/ARM64 jobs; publication was skipped. The downloaded quality and installed-runtime evidence archives match GitHub's recorded SHA-256 digests. The quality archive records 2665 passing tests, five skips and 32 successful plugin certificates with 405 passing results; all 1347 certificate evidence references match their hashes. Each platform archive contains 13 successful OpenTUI reports (88 checks) and nine successful runtime smoke reports. The distributed binary archives have not been independently downloaded for this audit, and the run predates later resource-monitor and navigation changes.

The build-only run for 876e2d1 passed verification and all four Linux/macOS x64/ARM64 jobs; publication was skipped. The downloaded quality and installed-runtime evidence archives match GitHub's recorded SHA-256 digests. The quality archive records 2658 passing tests, five skips, and 32 successful plugin certificates with 405 passing results; all 1347 certificate evidence references match their recorded hashes. Each platform's installed-runtime archive contains 13 successful OpenTUI reports (88 recorded checks) and nine successful runtime smoke reports. These are retained reports from the workflow, not an independent local verification of the distributed binary archives. They qualify neither later changes nor physical-terminal coverage or overall 1.0 readiness.

An additional local Qwen3 1.7B checkout diagnostic with the adjacent-duplicate edit advisory exhausted all 12 admitted requests and failed the independent behavior check. The package file survived, but the model edited the protected test fixture. This trajectory did not trigger the advisory, so it gives no evidence that the advisory improves model recovery. The separate deterministic replay verifies the earlier duplicated-line edit is reported without blocking it; it does not establish autonomous development. The isolated server exited; the retained report is dist/local-cpu-checkout-adjacent-warning-20261115/.

Source CI for 3f944f8 failed its Node 22 unit lane while waiting for a pending question in the OpenTUI host navigation test. The captured log shows the question was not observed within the poll deadline; it does not establish whether the cause was scheduling or a product defect. This failure remains open pending investigation. Other platform or local passes do not turn it into a pass. The subsequent b6ee298 Node 22 CI lane passed. A separate local Node 22.19.0 full suite passed 2635 tests with 34 skips, and the failing test passed 21 repetitions both normally and with two competing workers on one CPU. These observations did not reproduce or explain the original failure. The test now retains submission, request, core and recent event state in its failure message without changing the question assertion or its deadline. A separate injected missing-question fixture confirmed that this message identifies a completed submission, the received request and the relevant events; it does not explain the original uninstrumented failure.

The Node 26 lane of CI for a120abb failed its delegated-child visual check with an unterminated JSON record while reading a concurrently written trace. The exact bytes at the instant of the failed read were not retained. The observer now uses the existing complete-line parser, which waits for a terminating newline and still rejects malformed complete records. This addresses a possible partial-write race; the failed run remains failed and does not establish the exact write ordering.

At source revision 5133d5b, the configured OpenTUI CLI passed nine browser-backed PTY navigation checks and eight conversation-copy checks. These include keyboard retiling, narrow-window layout, closed Main recovery, pending questions and approvals, shared drafts, and saved layout after a real restart. The reports are in dist/wm-navigation-5133d5b/ and dist/wm-copies-5133d5b/; the latter covers native and selected-controller compositions separately. The PTYs exited successfully and relevant captures were inspected. These use scripted providers on Linux; they do not establish daily usability, other terminal protocols or exact-head release qualification. The earlier build-only run for 3ba17ac completed all four native build jobs; publication was skipped. Its result does not qualify the later source revision.

An isolated CPU run of the known recover-and-preserve-contract task at 08db772 failed after seven local requests: Qwen3 1.7B invented a source path, repeated searches in that missing path, and made no repair. The independent original checks and behavioral grader both failed; the protected package file survived. The owned local server exited and reported zero VRAM. This is one known-case diagnostic, not a representative model-family score. The repeated failure guard stopped the loop, but incorrectly attributed the errors to permissions or configuration; the guard now directs users to the actual tool errors without assuming their cause. The full development gate remains open. In a separate run at 2b78369, the same known case and independent checks used the explicit @. project-directory reference, changing the prompt and exposed context. The model found and edited greeting.mjs, but the edit still failed the original and behavioral checks; it ran npm test with exit code 1 and exhausted the twelve-request limit. The package file remained unchanged and the owned CPU server exited. This assisted run does not demonstrate autonomous discovery or a correct repair; its changed prompt prevents a like-for-like comparison.

An exploratory serverless CPU onboarding diagnostic now runs a pinned local llama-completion executable directly, with no model server or product actions. On twelve known intent cases, the 0.6B model matched five and the cached 1.7B model matched ten; both followed injected instructions in two cases. Valid JSON did not imply a trustworthy decision. See the diagnostic and its limits. This does not implement or qualify embedded-model onboarding. In five additional synthetic dialogue situations, both models returned valid JSON but falsely claimed that plugins were enabled when prompted to do so. The diagnostic exposed no action capability; its text still requires human review. These observations rule out treating structured output as sufficient for truthful onboarding.

Build-only run 43687179751 for 85ac6dc failed its macOS ARM command cancellation scenario before the first cancellation check: the captured input clicked row 20, while the retained failure screen showed the answer button on row 40. The exact intermediate frame sequence is unavailable. The visual bench now requires the complete terminal contents and geometry to remain unchanged for 150 ms before its single click. A source PTY run passes all three original cancellation checks; this does not make the failed archive run pass or establish macOS behavior. A fresh native run remains necessary.

Source checkpoint 235abdd passes CI 43647694838, documentation 43647694790 and plugin preparation 43647694862. Actual downloaded artifacts contain 2526 passing tests plus five optional skips on Node 26, 2515 plus sixteen skips on Node 22, 88 source UI checks, 20 successful PTY exits and 32 plugin certificates. Its installed-CLI synthetic Responses cases separately prove natural-exit drain and cancellation accounting, with zero external HTTP; their failed plugin grades remain truthful because the scripted actor delivered no plugin.

The installed 2ca0b69 runtime additionally passes a finite six-check Linux/Kitty complex-grapheme and real OS-window resize flow. Native cursor positions, Backspace/undo, preserved drafts, submission, 852 unchanged input bytes and natural PTY/TTY cleanup are verified. The first failed attempt forwarded a stale size in its relay; it remains failed. The corrected relay observes the public TTY resize event and preserves the original business expectations.

The following media repair adds compact producer status, visibility and clock recovery, immediate companion controls, cancellation fences and capability-based Kitty/SIXEL selection. Its 90 focused source tests include actual SDK rendering, modern and legacy delayed-media delivery and hidden-pixel disposal. Original assets are unchanged. A new archive and physical all-state/compositing checks remain separate requirements; this source result does not close those gates.

The first full media run passes 2553 tests with five optional skips and fails one public-render test's expected dispatch count. A controlled real-host trace reproduces the fourth dispatch when a pending layout save completes during the fake-clock test; the physical painter limit remains two and stale output is rejected. The original full run has no scheduling trace, so its precise ordering is not established. The test now drains actual layout persistence before measuring explicit invalidations, retaining all capacity, authority and final-publication assertions. Fresh standalone full verification passes 2554 tests plus five optional skips in 276 files, with no unhandled errors and all 1030 tracked/nonignored source files unchanged during the run.

The resulting bf5e61b media candidate is not qualified. Its sealed Linux archive passes nine canonical smokes, including six signal cases, but physical Kitty verification stops at the initial idle cycles with zero of fourteen checks complete. Frames arrive roughly every 750 ms and skip source positions. The exact-head Node 26 unit lane also passes 2554 tests, while its subsequent Ghost PTY check fails the unchanged original-frame timing oracle. The existing installation is retained; neither result is reclassified by other passing jobs.

A focused CPU profile identifies React development tracing recursively expanding the 207936-byte RGBA buffer passed as a component prop. An accessor-only control keeps the same pixels, FRAME listener and source clock and restores the preview cycles. The maintained repair passes 36 focused renderer tests and typecheck; the ordinary source Ghost PTY command, without a diagnostic loader, measures idle 1084 ms and thinking 1085 ms against the unchanged 1080±180 ms bounds. A deterministic SDK regression observes real development profiling, rejects the old buffer-prop expansion and verifies two fresh native RGBA/alpha frames. Fresh exclusive full verification passes 2555 tests plus five optional skips in 276 files, with all 1030 source inputs unchanged. A distinct sealed archive's physical checks remain required. This does not establish all-state, terminal or stable 1.0 readiness.

Retained source and archive history​

The local 2ca0b69 checkpoint passes the full standalone suite: 2386 tests plus three optional skips in 266 files, with no unhandled errors. Whole-workspace types, generation, documentation, marketplace validation and all 32 mandatory plugin contracts pass. Actual source CLI checks cover graceful SIGINT/SIGTERM, owned SDK/native process reaping, unrelated process preservation, long ignored hooks, concurrent close joining, journal failure and truthful final results. The new canonical installed-runtime gate adds six signal cases as a ninth smoke report. The fresh committed 1c5c3a8 archive passes all nine canonical smokes, including the six installed signal cases. Its native run passes 72 checks in eleven runners, then fails the public-render limit: eight actual held painters remain unresolved before any fixture gate is released, despite the limit of two. The remaining wire and delegated-child runners were not executed. This archive is retained as failed evidence and has not replaced the installed binary. A narrow dispatch/SDK-authority separation is committed as 2ca0b69 and passes 27 focused host/cache tests, all eight actual source PTY checks and a fresh sealed archive's full 95 native checks in fourteen runners and twenty-two PTYs. All nine canonical archive smokes pass. Captured continuations and implicit stale fallthrough are guarded; current dispatch results and publication remain verified. The exact build-only release 43635594880 passes verification and all four native jobs, with publication skipped. Its four downloaded, digest-verified artifacts contain 352 UI checks, 36 smokes, 24 signal cases, 32 reaped owned workers and 80 successful PTY receipts. This archive is installed, with the prior installation retained and Ink still the default.

Overall qualification for 2ca0b69 remains false: source CI 43635573480 fails the same eval-transport request-cap test on Node 22 and 26. The response can reach a client before receipt persistence releases admission; a real SSE client also exposes a mistaken upstream cancellation after consuming [DONE]. These are separate from the passing packaged UI/runtime evidence. The bounded serialization and terminal-reader correction passes 24 transport checks, eight SDK checks and the canonical full local suite: 2402 tests plus three optional skips in 266 maintained test files, with no unhandled errors. Retained dist/ snapshots are excluded from discovery. The correction also closes never-ending incoming bodies on cancellation, checks client loss before upstream admission and joins concurrent close calls. At that tooling checkpoint all 735 conservative bundle rows remain unchanged. No failed CI is retried or retroactively passed; later corrections receive their own exact-head qualification.

The source-only correction is pushed as e862952 and 15675d6. Both Node jobs now pass all 24 transport cases, while source CI 43641127303 fails two different UI test observations: Node 26 does not observe the initial Main question within a short poll, and Node 22 checks the final painter before awaiting its actual outer dispatch. Documentation and plugin preparation pass. Test-only corrections now wait for actual question publication and the exact current render completion; all original state, privacy, capacity and ownership assertions remain. The exact slow stage of the failed history run is not established by its retained artifacts. These test corrections pass the full standalone local suite: 2402 tests plus three optional skips in 266 files, with no unhandled errors. Their committed e887afb source CI 43643133938, documentation 43643133917 and plugin preparation 43643133980 all succeed. Downloaded, digest-verified artifacts contain Node 26's 2402 passing tests plus three optional skips, Node 22's 2391 plus fourteen skips, 88 source UI checks, 20 successful PTY exits and 32 plugin certificates. Packaged runtime code is unchanged; older failed workflows retain their original results and the installed archive's native evidence is not rebound to this later source.

The saved-profile development launcher now uses native Responses with one parent-owned reservation/collector/settlement authority shared by the typed evaluator. Development reuse requires the original existing lifetime ledger; preparation reads no saved profile, credential or ledger. Focused fixtures cover raw reasoning/function continuity, bounded admission, durable evidence and cancellation. A separately selected actual installed executable passes natural exit before upstream EOF and operator cancellation during the final drain, using a synthetic profile and zero external HTTP. That scripted actor delivers no plugin and its authoring grade correctly fails. These checks establish the transport/lifecycle boundary; repeated independently graded real development and initial useful-plugin delivery remain open. See the launcher contract.

OpenTUI runs the configured application through --renderer opentui; Ink remains the default. Revision 3281bda adds private delegated and managed agent request windows, genuine producer/controller ownership and source-Core lifetimes. Independent managed requests survive Main resume; expired Main descendants are revoked. These behaviors pass ten actual-CLI checks, including selected-controller leases, privacy, exact effects and managed archive/shutdown.

The sealed Linux archive passes 95 native checks in fourteen runners and twenty-two PTYs, eight canonical installed-runtime smokes and independent source, shipping-byte, receipt and visual reviews. All 5879 shipping entries match the relocated archive. Final standalone coverage passes 2328 tests with three optional skips in 257 files; types, documentation and all 32 PR plugin certificates pass. The earlier local installer extraction failure from a full memory-backed temporary directory is preserved; the unchanged payload passes using disk temporary storage.

The exact 3281bda build-only release 43616152753 succeeds on all four platforms. Its downloaded, digest-verified artifacts contain all 352 OpenTUI checks, 32 installed-runtime smokes and 80 successful PTY receipts. Global qualification still fails because its Node 26 source CI clicked transient reopened-pane geometry. The subsequent bench-only 1ff271a requires the complete normal footer, focused owner and stable full-frame geometry before its single click. Its fresh source CI passes the pointer scenario but fails a later one-shot startup-trace read after six of nine navigation checks: the ready file is visible before the awaited SDK call returns and appends its completion receipt. The preserved final trace proves that exact phase completed without a fault. Documentation and plugin preparation pass. Its build-only four-platform release 43618662809 fails, with a separate Linux x64 appearance cleanup failure after all eight feature assertions: the final Ctrl+C interrupts a still-finalizing turn. The actual journal records that interruption; the pre-exit screenshot still shows Working and Stop.

The current narrow bench repair observes the exact completed startup phase using the existing 20-second bound and original oracle. Fresh source and sealed-archive navigation each pass all nine checks, with the actual Production answer preserved through close/reopen and restart. The runtime inputs, executable and renderer are unchanged. A fresh exact-head CI/platform qualification is still required; neither failed source CI is retroactively passed. This repair is committed and pushed as ed2d2ba; its fresh source CI, documentation and plugin preparation pass. The appearance bench repair waits for the exact restored turn's completion and painted Ready state before its single exit key, preserving the original feature and exit oracles. Fresh source and exact sealed-archive appearance each pass all eight checks, both native and selected compositions exiting cleanly. A new exact-head native matrix is required.

Actual isolated Alacritty/Ptyxis probes establish genuine terminal replies, PID-owned native screenshot capture and clean TTY/process restoration. Neither configuration negotiates Kitty graphics or advertises SIXEL. A separate actual installed-archive run in isolated Kitty 0.50.0 emits two original six-frame loops each for idle and thinking, with genuine terminal replies, exact uploaded RGBA including alpha, native captures and clean exit. Accepted terminal-write timing is 1075.79/1084.98 ms and 1083.66/1075.06 ms against the declared 1080±180 ms bound. The two native captures show a stable original frame in each state; this does not measure every GPU-visible animation frame or screen frame rate. The terminal's modifyOtherKeys advisory and mode 2027 warnings remain preserved compatibility issues. A separate installed 2ca0b69 run in a private rootful Xwayland/Kitty environment passes all four declared oracles: real unmodified CJK, emoji and combining keys, exact UTF-16/cell positions and edit/backspace/undo, canonical Tools opening and Main draft/cursor restoration, and exact prompt submission with natural PTY exit and TTY restoration. All six native captures and raw input/forwarding bytes are retained. Earlier helper failures remain preserved, including the incomplete ANSI observer that did not handle Kitty's OSC 66 width metadata. The successful flow retains 15 unsupported mode 2027 messages and one modifyOtherKeys advisory without filtering. This text/key path does not establish graphics or mode 2027 compatibility. A later source-tree OpenTUI run in Kitty 0.50.0 and tmux 3.8 with allow-passthrough on exercised the actual Ghost composer: changing frames, pause/resume, off/on, resized draft restoration and normal exit passed. With passthrough disabled it retained the compact status. This is one Linux terminal configuration, not installed-archive or cross-platform certification. The same physical source flow passed again at 876e2d1, with normal exit and private display cleanup in both passthrough modes. Captures and reports are in dist/tmux-research-20261115/app-tmux-source-876e2d1/ and dist/tmux-research-20261115/app-tmux-disabled-876e2d1/. Separately, an emulated Kitty PTY observed original idle and thinking frame cycles of 1082 ms and 1081 ms; this is not a physical screen-frame-rate measurement. The actual application appearance scenario also passed on source with a scripted provider. Full state transitions, physical compositing, resize, fallbacks, broad multiplexer compatibility and video remain open, as do the default frontend switch, CPU-assisted onboarding, repeated realistic development, initial useful-plugin delivery, sustained sessions and trust remain open gates.

An additional source-tree plugin-raster check in Kitty 0.50.0 and tmux 3.8 observed the fixture's red pixels with passthrough enabled, and its alt text with no red pixels when disabled. Both processes exited normally and the private display was cleaned up. Component tests separately exercise frame replacement and image deletion on hiding. This is not installed-archive, resize, screen or cross-terminal coverage. Evidence is retained in dist/tmux-research-20261115/plugin-raster-20261115-on-v2/ and dist/tmux-research-20261115/plugin-raster-20261115-off/. At 9802f1b, a further source run used an isolated plugin and actual terminal input to replace a red raster with green, resize from 110×36 to 80×24 and back while retaining a Main draft, close the chart in Arrange mode, and reopen it through the window finder. Screen captures show the old pixels disappear on close and the current image return on reopening. The app exited normally and the private display was cleaned up. The 80×24 and restored checks waited for matching Kitty and tmux geometry and complete application frames. Evidence is in dist/tmux-research-20261115/plugin-raster-window-close-9802f1b-v3/. Two earlier close attempts timed out on the test's focus observation; the second capture shows the focused narrow title was truncated. This is one terminal and synthetic plugin, not proof of other terminals or packaged-archive behavior.

The resources monitor now offers an opt-in raster plot alongside its text chart and retains a plugin-lifetime handle for periodic updates after the opening action finishes. A source OpenTUI/Kitty/tmux check observed successive samples, keyboard activation, the graph after PageUp and stable scroll during updates; passthrough-off separately showed the text alternative. The initial unfixed attempt failed after its captured foreground UI context expired. These are source checks in one terminal, not native-archive certification. Evidence is under dist/tmux-research-20261115/resources-graphics-real-final-on/ and dist/tmux-research-20261115/resources-graphics-fallback-off/. The later source check confirms OpenTUI reveals the newly enabled chart or its fallback without PageUp. A gated independent plugin scenario also confirms that manual PageDown while the image is pending is not reset to the top when it arrives. Reports and captures are in dist/tmux-research-20261115/resources-auto-reveal-final-on/, dist/tmux-research-20261115/resources-auto-reveal-final-off/ and dist/tmux-research-20261115/surface-raster-scroll-intent-v2/.

Plugin raster protocol selection also has SDK-renderer coverage for SIXEL with and without pixel geometry, including returning to alt text when geometry is lost. A real Kitty/tmux source regression separately rechecked image update, resize, window close and reopen after that change. On 2026-11-15, a separate source-tree check at f883ab1 launched OpenTUI in XTerm 411 configured as a VT340 with AGENT_GRAPHICS=sixel. An isolated synthetic plugin displayed its red raster and text in the sidebar, and the application exited normally. A follow-up synthetic plugin replaced the raster with green, removed its view and recreated it with red; screen captures showed the old colored pixels disappearing at each transition. Resizing the XTerm window down and back restored the chart; an intermediate screenshot briefly contained both old and new image positions before the final redraw. A control run with AGENT_GRAPHICS=off also captured duplicated text and panel borders immediately after the same resize, so the artifact is not limited to the SIXEL image path; its underlying cause remains unresolved. An experimental reduction of OpenTUI's resize debounce from 100 ms to 20 ms produced earlier settled captures in three XTerm trials, but did not eliminate the intermediate artifact. In two additional synthetic trials of 75 resize requests spaced 40 ms apart, the Bun process used about 1.92–1.93 CPU seconds over three seconds at 20 ms, versus 0.05 seconds at 100 ms. Both restored the chart. Two trials at 50 ms used 0.06 seconds in this same 40 ms-spaced scenario, but do not establish behavior at slower resize frequencies. This narrow XTerm/SIXEL fixture is not a general benchmark; the 20 ms change was reverted pending a better balance of responsiveness and resource use. A fresh source-tree control at 9b8a750 reproduced the text-only artifact in the same isolated XTerm fixture with AGENT_GRAPHICS=off: immediately after expanding the window, the first capture shows duplicated workspace text and an old panel edge; a later capture shows the settled layout. The PTY exited normally and the private display was removed. The report and captures are in dist/sixel-xterm-research/text-resize-current-9b8a750/; the retained local probe has SHA-256 13532d44b36475d5b66110f70225b83407754928d7f626cbba5255b0846b981b. This confirms a visible intermediate artifact in one environment, not its cause or a regression across all terminals. A blind reduction of the debounce remains insufficient evidence of a fix. An initial purported 50 ms trial changed the unused wire renderer, so it is not evidence about the application's debounce. A subsequent single trial changed the native renderer actually used by the application: its first restored-frame capture still shows the artifact. That temporary change was reverted. Captures are in dist/sixel-xterm-research/text-resize-native-50ms/. This is not a CPU or terminal compatibility benchmark. Another temporary native-renderer experiment called its immediate resize API on the first SIGWINCH, throttled to at most one call per 120 ms. Instrumentation recorded both 87×28 and 110×36 callbacks; the first post-expansion capture still showed the artifact and the next was settled. The first capture began in the same millisecond as the expansion callback, so it does not establish the relative ordering of the terminal's paint and the renderer output. This experiment was also reverted; screenshots, callback times and capture intervals are in dist/sixel-xterm-research/text-resize-native-leading-timed/. Its script and temporary instrumented renderer are local diagnostic evidence, not a shipped fix or proof of bounded CPU under continuous resize. Evidence is under dist/sixel-xterm-research/timing-* and cpu-*. A further run focused the panel, closed it with its mouse [x] control and reopened it through the window finder; the screenshot showed no colored pixels after closing and the image returned after reopening. A reference SIXEL image also rendered in the same isolated X11 environment. Reports and captures are under dist/sixel-xterm-research/. These checks do not establish absence of transient resize artifacts, other terminal implementations, tmux passthrough, or packaged-archive behavior.

Plugin development and interface usability​

On 2026-11-16 the combined OpenTUI manager scenario passed on source and on a locally relocated Linux x64 archive. While Main waited for a question, it searched the local marketplace, installed and enabled a plugin, saved a window placement and completed the answer. It then closed and reopened Main using Finder, restarted the CLI and recovered both the conversation and an unsent draft without a model request. Reports and screenshots are in dist/daily-flow-integrated-retry-20261116/ and dist/daily-flow-integrated-native-20261116/. This loopback PTY check does not establish a physical-terminal experience or certify other platforms. An extended run on source and the same relocated archive additionally refreshed the isolated catalog, detected version 1.1.0, queued the update while Main was busy, verified version 1.0.0 remained installed until settlement, and confirmed that version 1.1.0 remained enabled and loaded after restart. Reports are in dist/daily-update-source-verify-20261116/ and dist/daily-update-native-final-20261116/.

The 1.0 objective explicitly includes developing missing plugins needed for everyday project work and improving the terminal interface until complete user workflows are convenient. This is implementation scope, not only an evaluation of existing features.

  • Identify missing capabilities from real development tasks. Extend an existing plugin or create a focused plugin through the public extension API; avoid adding host-specific integrations when an extension can own the behavior.
  • Exercise the complete plugin lifecycle: discover, install, enable, configure, use, update, disable and remove. Verify failures, recovery, reload and saved settings as well as the successful path.
  • Make plugin search, tools and settings usable while an agent is running. Preserve drafts and task progress while navigating, answering a question or approval, cancelling work and returning to the conversation.
  • Review keyboard and mouse operation, focus, visible progress and errors, narrow terminals and restart persistence. Define concrete scenario criteria before changing the interface; validate them in the real PTY bench using the project's tui-visual-check skill.
  • Use Compositor itself for plugin and UI development where practical, then independently review the patch, execute behavior checks and inspect actual terminal screenshots. A model's success report or passing headless tests cannot establish interface usability.

Prioritize gaps that block daily work, then repeated friction observed during use. This scope does not raise the authorized live-model spending limit.

An isolated XTerm 411 check of the default Ink frontend exposed a light-terminal contrast issue: with XTerm's default white background, much of the setup guide was very pale and hard to read. Ink now uses a light default palette when a valid terminal background reply indicates a light background. Follow-up captures for white and dark backgrounds are under dist/sixel-xterm-research/setup-corrupt-ink/ and setup-corrupt-ink-dark/. This source check does not establish behavior for every light terminal, missing background replies, or custom plugin palettes. Intermediate and saturated RGB backgrounds now receive adjusted default text colors, with contrast checks against the reported RGB value. A gray XTerm using its limited ANSI palette mapped bold black text to low-contrast gray. On light 16-color terminals Ink now substitutes underlining for bold black; the same XTerm fixture showed a readable heading. Other ANSI palettes and custom themes remain separate compatibility cases.

Accepted usability work — 2026-11-13​

The user's daily-use feedback adds these implementation requirements:

  • Approval review must handle concrete CI/shell/plugin actions with the user's actual task context and a validated structured decision. Enabling auto should not silently require a second model profile. Explicit denies stay binding; malformed or unavailable review is distinguishable from an actual decision. Ordinary clarification questions must remain separate from permission review.
  • Commands need an immediate, non-model execution path independent of whether they render a view. /auto should not require an otherwise unhelpful panel. Long operations need progress, cancellation and an optional deeper view.
  • First launch should explain profiles and available plugins, recommend a useful starting selection and help apply the user's choices. Investigate an optional small CPU model embedded without an Ollama/server requirement. No implementation or quality claim for that embedded model is established yet; first-run setup must remain usable when no model is available.
  • Workflows and agents need reusable interactive tables, expandable details, tabs and drill-down into an individual agent's actions. Replace static row/column dumps where those interactions improve navigation; preserve keyboard/mouse operation and narrow-terminal behavior.
  • Provide a convenient side conversation such as /btw with an independent agent while the main task continues. A closed main conversation must be easy to restore without losing its process, history or draft. Existing Alt+Enter conversation creation needs explicit discoverability and behavioral proof.
  • Support useful views that need no agent, including live resource charts. An explicit action may share a selected observation with an agent. Task-service integrations such as Jira are examples to evaluate, not authorization to connect an account or create external tasks automatically.

These are acceptance requirements, not a claim that the current release supplies them. Prioritize approval behavior and conversation recovery, then the common command/view contracts and onboarding. Deterministic checks do not substitute for independent review of the complete real terminal flows.

Version 0.8.0 ships catalog 0.25 immediate commands independently of views, headless /auto on|off|status, an explicitly requested /auto history view and idempotent /chat restoration. Real PTY checks cover a held model, unchanged draft, repeated restoration, workspace return, narrow history and restart. Core checks cover owner dispatch, failure isolation, tool permissions, auxiliary usage, cancellation, unload and shutdown bounds.

The bundled resources plugin provides a standalone non-model window with live process CPU/RSS and OS-reported system CPU/non-free memory charts, pause/manual sampling and bounded in-memory history. System counters may describe the host rather than a container's quota, and non-free memory can include cache. The process chart includes in-process agents. On Linux, a separate child CPU chart estimates the load of surviving descendants and an on-demand snapshot sums their RSS; both report incomplete observations and can miss short-lived processes. Separate attribution for external model servers, per-agent attribution and broader resource integrations remain open. An explicit panel action can queue the latest complete selected sample for the current agent; sampling and opening the panel do not request a model. Agent and workflow views use catalog 0.27 semantic tables and catalog 0.28 tabs: expandable details, state filters, paginated stages/results, keyboard/mouse navigation and narrow stacked fields. Workflow pause/resume and legacy journals remain available. Closing the workflow view leaves stages running, rejects stale renders and allows explicit reopening.

Per-agent conversation tabs preserve separate drafts and histories, provide button/Enter submission and navigation through all sixteen allowed agents, and restore the selected pane with committed messages after restart. Agent Reply drafts are stored privately and survive restart after the debounced save; the main editor and private form drafts are not persisted. A source OpenTUI PTY scenario verifies fast switching and restoration of two Reply fields after process replacement, and sends one reply through an isolated loopback provider. Conflicting later input is retained in-session for explicit restoration or discard in both frontends. This does not establish external model quality. Saved activity exposes persisted prompts, turn lifecycle/usage and tool arguments/results/errors through filtered, paginated expandable records read through the configured session repository. Linux PTY checks cover an actual read tool call, a held main stream and restored saved output. Full native UI coverage on other platforms and actual terminal graphics remain separate gates from native installer/protocol checks.

/btw opens an independent persisted agent with its own Reply field, immediate reopening, separate draft, focused-agent cancellation and restart history. /btw --once preserves notification-only behavior. Both modes ship in 0.8.0; the bundled btw package is version 2.0.0. Source integrations and compiled Linux PTY scenarios cover these flows while a main model is held.

Auto-review validates structured approved: true|false|null decisions. Task-context low/medium approvals can proceed; refusals stop the call without another manual question; uncertainty and review failures remain manual. Explicit host denies remain binding. Missing targets/effects require uncertainty rather than a fabricated violation. Full Linux source release-stage certification on exact release source 4b5b297 passes the declared gemma4:31b-cloud matrices for auto-approve 1.4.1, workflows 1.6.1, btw 2.0.0 and plugin-author 1.4.0: 24/21/24/21 cases, unchanged oracles and three repetitions. This is one-reference evidence. Ordinary clarification questions remain human decisions; arbitrary model classification quality is not established.

The plain/headless interface preserves independent controls in pasted input, queues Enter follow-ups, reserves Tab for steering, drains accepted commands on EOF and consumes human dialog answers separately from model prompts. Command notifications and auxiliary usage appear in headless results; command errors fail that invocation. Installed-runtime checks cover actual raw PTY approval, interruption and CLI exit on all four declared platforms.

The optional first-launch /setup guide requires no configured model. It offers a short private provider editor, suggests explicit optional plugin choices, preserves installed pins, applies reload through the normal UI path and returns to the first task without losing a draft. Eleven compiled Linux PTY checks cover the automatic offer, persisted skip/completion, masked/private credentials, selected-plugin installation/activation, narrow layout and one first task through a saved loopback connection. Embedded CPU inference-assisted onboarding and its model quality remain open.

0.8.0 publication — 2026-11-13​

v0.8.0 is published from 4b5b2972b23b0a8c20c813f076b14ba601d77849. Candidate verification and tagged publication passed source checks and native build/install jobs on Linux x64/ARM64 and macOS x64/ARM64. All 32 tagged installed-runtime reports were independently audited. Matching Node 22/26 CI passed; coverage passes 1315 tests with three optional skips. All 4487 Linux candidate shipping files match the locally exercised archive, including executable bits and symlink targets. Real setup, conversation-tab and workflow PTY scenarios pass with inspected narrow/restored/completed screenshots. All six public assets were downloaded; the four archives and installer match published checksums and all available GitHub asset digests match the downloaded bytes. The public Linux distribution has the same 4487 shipping files as the verified candidate. An isolated public 0.7.4→public 0.8.0 installer upgrade preserves settings, plugin state, draft, old executable and the old session's journal prefix; the new host resumes its original task/answer. Versioned and latest installs both pass. The upgrade uses two bounded loopback completions, not a live provider. Public integrity evidence is retained under dist/release-0.8.0-public. This intermediate release does not close the stable 1.0 gates.

Two-family coding diagnostic — 2026-11-13​

The unchanged two-task coding suite ran twice each on gemma4:31b-cloud and DeepSeek-v4.1-flash:1017-cloud through the existing Ollama service, using the 0.8.0 runtime code. Both score 3/4, for 6/8 overall. One Gemma run repairs the artifact but repeats a denied scratch-test command instead of running the existing allowed project checks. One DeepSeek run passes all artifact/original-check oracles but cannot admit its next request within the declared 80000-token turn allowance before terminal completion. Neither failure is reclassified as success. Trace/report/task/runner/grader hashes are retained under dist/evals-v080-cross-model.

These are known small tasks, not held-out general development benchmarks or a broad model-independence claim. The original criteria and limits were unchanged. No OpenAI profile/request or spending-guard change was used; Ollama monetary billing is unavailable. Denial recovery, efficient completion and broader real project/self-development coverage remain open.

Follow-up recovery experiments did not establish a fix. A supplemental general instruction scored 7/8. A longer model-facing denial footer scored 8/8 in the ordinary suite, but its successful Gemma greeting runs never encountered a denial. A separately declared controlled probe inserted one real denied command after a successful live edit and before verification, then used live models for the remaining work. Across two repeats per family, the prior footer scored 2/4 and the candidate 1/4 under the stricter recovery criteria (one controlled denial, no further denials/questions and successful original project oracles). These small exploratory results do not establish a general performance difference; they did not justify promotion. The candidate footer was withdrawn. Original tasks, graders, permissions and budgets were not weakened. Evidence and the withdrawn source hash are retained in dist/verification-recovery/decision.json. The next investigation is existing check discovery and task context rather than another unverified instruction-only change.

Finite-query repetition recovery — 2026-11-13​

The default loop now warns after three repeated query-only steps and stops after six, using optional ToolSpec.loopGuard: "query" metadata. Built-in read and Callix glob/grep opt in; polling remains unchanged. New observations, other operations, steering and context replacement permit fresh investigation. The recent window is bounded; see exact rules.

The 0.8.0 tagged native matrix passes the query probe on all four platforms, including real read/grep journal evidence for bounded refusal and recovery after a new observation. The earlier archive failed the new probe. Final coverage passes 1315 tests with three optional skips.

Two prototype CPU greeting attempts still score 0/2 with 17 admitted requests. Neither reaches the reminder threshold, so this is not a measured recovery or efficiency gain. The final context-replacement reset is covered by regressions and the rebuilt archive, not by that earlier compaction-disabled model run. Novel-but-unproductive queries and proposal-only responses remain separate gaps.

No-change edit evidence — 2026-11-13​

The default file-edit resolver rejects a replacement whose actual resulting text is unchanged, before writing or adding undo state. Useful whitespace normalization still works when identical arguments change the actual file. Five regression cases cover exact/normalized single/all replacements and an empty guarded insertion. The six-response discovery/read/no-change-edit probe passes on all four platforms in the 0.8.0 tagged matrix; the earlier archive fails its no-change check. These establish tool behavior, not autonomous model correctness.

Two subsequent known greeting diagnostics still score 0/2 with 21 admitted local requests. One guarded edit introduces invalid syntax and fails completion checks; the other returns a proposal without implementing it and fails independent grading. Neither invokes a no-change edit, so the model runs do not directly validate its new rejection or establish improved model quality. The native and unit regressions provide the rejection evidence.

Ordinary directory-reference coding diagnostic — 2026-11-13​

The CPU evaluator can now exercise the existing @. directory-reference flow using --project-reference. Under this explicitly different context condition, Qwen scored 1/2 on the known checkout task and 0/2 on the known greeting task. The successful attempt performed a guarded edit, ran the real project checks and passed independent behavior checks while preserving protected files. Across the four attempts, 42 local completions were admitted; no paid request occurred.

Those historical failures exposed repeated successful read/search loops and an identical-text edit reported as successful, followed by failing verification. Completion remained failed. The later query/no-change protections have separate regressions; these original runs do not prove that directory context alone supplies reliable coding. Prompts, source hashes, budgets, failures and cleanup are retained; this is not an unseen benchmark or an unchanged-prompt comparison. See local evaluation controls.

Current verification evidence after reruns — 2026-11-13​

Build-only run 43320159884 at 7139daa passed verification and all four native build/install jobs. All 28 downloaded runtime reports pass, including four coding-completion scenarios on each platform with actual process exit codes and run counts. Publication was skipped. This establishes the runtime checkpoint, not a public release or broad real-model readiness.

The coding completion policy no longer accepts an earlier successful check when a subsequent attempt fails. Every recognized check attempted after the latest built-in edit requires current passing evidence; another passing command does not hide a failure. Actual process tests and the installed Linux archive prove refusal after pass/fail and recovery after a third successful check. The native release probe now requires all four unchecked/checked/regressed/recovered cases. The same four scenarios pass on all platforms in the published 0.8.0 matrix. These policy checks do not establish live model quality or stable 1.0 readiness.

The local CPU evaluator can explicitly compose read/write/edit/exec with Callix glob/grep and records the tools actually transmitted to the model. The known multi-file coding diagnostic still scores 0/2 both before and after clearer discovery/read guidance; admitted calls increase from 14 to 24. This does not establish a model-quality improvement or readiness for autonomous development. The case, independent grader and request/context/output limits are unchanged.

Read now exposes its integer, 1-based pagination limits and rejects invalid values with specific recovery guidance. Empty file/text searches and invalid directories explain how to discover actual paths without inventing them. The six-response discovery/read/no-change-edit probe passes on every native platform in the 0.8.0 release. These checks establish bounded tool behavior, not general model competence. See evaluation guidance and release verification.

Workflow close boundary regression — 2026-11-13​

The prepared 0.8.0 source candidate exposed an asynchronous workflow render arriving after its view was closed. Render revisions now reject stale work and close waits for already-submitted publications. A deterministic regression and the required real PTY scenario verify background workers completing while the view stays closed, followed by explicit reopening. Source and compiled Linux scenarios pass. These are controlled checks; the published 0.8.0 matrix now supplies all four native build/install gates and final coverage of 1315 tests. They do not establish broad model quality or 1.0.

Follow-up Node 22 CI found premature acknowledgement when a background refresh superseded a navigation render. The queue now covers loading and publication and navigation waits for the replacement frame. A deterministic regression fails on the prior implementation. The fixed queue is included in 0.8.0 and passes the real PTY workflow scenario and final native matrix.

The refreshed local Linux archive also passes the real workflow PTY scenario and all eight installed-runtime probes. Its closed, completed and narrow screenshots were inspected. Source workflow certification passes all bindings and 21/21 live reference tasks on gemma4:31b-cloud. Generated workflows 1.6.1 ships in 0.8.0; public shipping files match the exercised Linux archive.

Planned OpenTUI migration — 2026-11-13​

The user explicitly requested adding migration to OpenTUI to the plan. OpenTUI is the intended future terminal frontend. The opt-in application frontend now runs the configured app with compositor --renderer opentui. The deterministic conversation and workflow preview remains a separate development fixture. Ink remains the default. The acceptance checks below govern the switch to a shipping frontend. Preserve the agent core, provider adapters, permissions, tools, session state and extension ownership contracts throughout the migration.

  1. Build a separate renderer-opentui prototype using the existing Bun/React stack. Exercise the message editor, the original Ghost beside the composer, agent tabs and one expandable workflow table. Verify actual interaction and media behavior before extending the implementation to all screens.
  2. Evolve the UI description where needed so OpenTUI receives reusable interactive components rather than only preformatted rows/cells. Keep existing plugin views usable through an adapter; plugins must not require a wholesale rewrite or execute arbitrary frontend code merely to show a view.
  3. Move the production conversation, input, approvals/questions, onboarding, settings, plugin management, workflows and agent drill-down onto the new frontend. Use native input/selection/scrolling components and explicit expandable details. Preserve the compositor's window-manager model, including the existing tiling tree, independent focus, workspaces, drag/drop, resizing, close/reopen and saved placement. Keep non-model commands independent of rendering.
  4. Verify the original character assets, transparency, frame timing, state transitions, resize/clipping and terminal protocols. Test actual APNG/GIF animation support rather than inferring it from static image decoding. Retain the user's choice against substituting block artwork. Video is an additional supported media path to evaluate, not a reason to convert transparent assets to an unsuitable format or invent animation frames.
  5. Pass real PTY keyboard/mouse/paste/Unicode/focus tests, draft and session restoration, concurrent tools/agents, and compiled archive/install checks on all declared platforms. Verify native library/resource packaging. Switch the default renderer only after the replacement passes those gates; retire the Ink frontend after the migration is complete.

The first prototype passed seven Linux browser-backed PTY checks at 100×30 and 48×18. A held streaming Main session remained active during a Side reply, workflow row expansion and draft recovery. Checks covered mouse selection, keyboard navigation, a multiline Unicode paste and process exit. A separate PTY check matched every emitted RGBA frame in the idle and thinking loops to the original Ghost images. The browser and byte-level checks emulate Kitty protocol support. Native terminal coverage, additional protocols, the full plugin view contract and packaged releases remain outstanding.

A follow-up Linux PTY run exercised concurrent permission requests from two agents. It verified per-chat badges, full argument inspection in a narrow terminal, switching to Workflows and returning with drafts intact, a one-time allow and a denial. Independent file checks verified that only the selected write took effect. A repeat requested permission again, and closing the preview while it waited preserved the preexisting file.

The source preview now uses the existing host-owned binary window manager. Native panes receive rectangles and retain their editor instances through nested splits, directional focus, workspace moves, drag/drop, resize, zoom and closure. The host persists the tree, ratios, selected workspace, focus and closed panes. Actual project installation, explicit enable/disable and deferred update/removal keep saved and loaded-version status separate. User placement-rule controls, provider/settings flows and selected controller adaptation still need migration. Production protocol binding, saved conversation recovery and release packaging remain outstanding.

Its main Linux PTY scenario passes 18 checks, including visible Main/Plugins panes, layout controls, independent approval routing, deferred updates, command-written version receipts before and after update, disable/removal and clean shutdown. The updated command exposed and then verified a fix for Bun retaining old package modules despite a changed manifest. These checks use a deterministic local provider and local upstream fixture, not a real-model development task.

The window-manager scenario also verifies close/reopen draft, cursor and undo state, actual title and separator gestures, workspace moves, compact projection and layout restoration after actual PTY restarts. An empty selected workspace must remain empty. These checks distinguish in-process editor retention from conversation recovery after process exit, which is not implemented in the preview.

A single-packet F5/Alt+Y input initially confirmed Main after host focus had already moved to Plugins. Native focus changes now commit synchronously, and action admission checks the current host snapshot. Real PTY bursts verify that hidden approvals stay pending, typed text reaches the newly selected chat without loss, and a following denial targets only that chat's request.

Native single/multiple-choice questions, custom answers, questionnaires and private text/password/URL/select/toggle forms now use per-core host callbacks. Each preview conversation arbitrates approvals, questions and forms in one FIFO. The input editor preserves an independent chat draft; a model can work while a private form is open. Foreground interruption cancels its question while a form owned by a separate user operation stays live. The source preview rejects selected custom question/form controllers without silently substituting a native view. Independent settings-pane placement and precise delegated-child request attribution remain production integration gates.

The inputs PTY scenario checks actual model-received answers and a trusted plugin's private result, rather than only UI labels. It caught and verified fixes for questionnaire text loss between pages and rapid Escape/mouse-release packets leaving a native pointer capture active. A synthetic password stays out of the terminal output, provider requests and session journals. The bench's raw keyboard log intentionally records synthetic test input; it is not a real credential test.

Initial technical references are the official OpenTUI components, image, video and runtime/platform requirements. The prior chat estimates were preliminary effort ranges, not a release deadline. This migration extends the 1.0 usability scope and does not change the authorized live-model spending limit or interrupt the current approval-work fixes.

Opt-in OpenTUI application integration — 2026-11-14​

--renderer opentui now connects the actual BuiltApp, provider router, private credential store, session repository and loaded plugin handlers to native conversation and semantic plugin windows. Explicit settings forms have their own operation and window, so Main's pending question does not block editing a connection. Loaded plugin views retain host-validated actions, revisions and ownership. The host adapts plain legacy text without executing plugin frontend code. The canonical placement editor, saved rules and full window tree are shared with the application. This closes the earlier preview-only integration gap; Ink remains the default.

The Linux source and relocated native application checks exercise real profile add/edit/cancel/discovery/selection, private credential and proxy storage, a pending Main question plus independent draft, foreground interruption without cancelling a separate form, and actual session/profile/layout restoration after restart. Requests use a bounded loopback HTTP provider through the ordinary router. No live-model endpoint or paid test budget is involved. The native archive carries the renderer bundle, native library, parser WASM and workers; release CI now requires the actual relocated application flow on all four declared platforms. Adding that gate does not establish that its first remote matrix has passed.

The production OpenTUI frontend also exposes stable Arrange and searchable Windows navigation over the same host window tree. Its native search and resident editors preserve independent draft/form state; the deterministic preview retains its separate development keymap. Verification must include actual batched keyboard/paste, protected questions/approvals, live route replacement and narrow terminal geometry, not only host parser tests. Arrange also supports Ctrl+Arrow edge retiling independently of Shift swaps. The nine-flow navigation gate includes a real three-window nested rearrangement, nearest-split resizing, no-neighbor behavior, pending Main/form preservation and exact restoration of the keyboard-created tree after process restart.

Selected approval, question and private-form controllers now use the shared host lease adapter and independent request FIFOs. The renderer receives validated private frames/cells rather than factory requests or native form defaults. Main keeps an explicit answer/draft input owner; selected approvals retain validated once/session/project results through the ordinary Core policy. This closes only the private-request portion of controller migration.

Selected editors now have a separate capability-bearing host adapter. Validated views and bounded receipts reach OpenTUI; command/file discovery and UTF-16 checkpoints stay private. Available focused Draft owns its plugin's F2 binding; the window toolbar and Alt+W then W remain reachable. Answer mode and other windows retain F2 Windows. Submit/steer use Core's actual admission receipt before editor input continues, preserving busy queues and restoring a visible draft only when ownership was rejected. Reload/resume retain state before disposal; old leases cannot submit to the replacement Core. This closes the selected editor implementation gate; actual source/native verification is recorded with the milestone evidence rather than inferred from host tests.

Selected history and placement implementations now use app-private windows. Canonical browse commands use the UI command route while Main is busy. History retains its semantic inputs, private forms, normal session service interception and selected repository; every input has fresh foreground provenance. Placement uses guarded windows/anchors/apply capabilities with target revalidation after save. Hidden state, cancellation and old-generation revocation are explicit. The eight-flow actual-CLI scenario passes from source and a relocated Linux native archive, including exact repository metadata, intercepted service calls, private results, restored context and saved layout. The same candidate also passes all 47 prior native checks. Real host regressions separately prove busy resume rejection, cancellation after Core preparation, same-owner closed-window and split-ratio retention, and current failure feedback after reload. New-head four-platform CI remains a required candidate gate; these local checks do not establish native terminal/protocol coverage or stable 1.0 readiness.

Saved Ink session, Tools and Help windows now migrate with their original IDs, closed state, serial ceiling, workspaces and split ratios. Main copies share one Core, native draft/cursor/undo, request FIFO and selected editor lease, with independent transcript scroll. Arrange n/t/? and Finder creation actions allocate explicit windows; /chat remains idempotent. The resident input stays mounted above opaque tile backgrounds and below Finder, including newly created panes. Every pane port is bound to its Core generation; expired ports cannot submit to a replacement, including a resumed journal with the same ID. Host windows are also valid placement neighbors while placement-rule targets remain plugins.

The eight-case actual Ink -> OpenTUI -> OpenTUI restart scenario passes from source and a relocated Linux native archive with native and selected controls, exact stored trees/identities, shared Unicode input, independent scroll, private form isolation, once-only callbacks, reload/fork/resume and a delayed old-generation submission. The same candidate passes all 56 earlier native checks. Exact-head four-platform CI remains a required candidate check. This implementation closes the missing legacy-window behavior; it does not establish native terminal/protocol coverage or stable 1.0 readiness.

Portable appearance now shares validated replacement/order/removal semantics with Ink while retaining OpenTUI's own defaults. A data-only, revisioned source updates native colors, frame chrome, role labels, welcome, optional hints and safe current-Main tool details without replacing input/controller owners. All ten color fields and named bright colors are mapped explicitly; selected rich cells retain their own declared colors. Title controls share six painted/hit-tested cells, with focus/hover/Arrange policy and narrow-window bounds. Explicit Tools visibility is publication intent rather than a continuously enforced layout. Historical tool groups with ambiguous UI provenance are omitted, and exact live call occurrences prevent provider-ID reuse from exposing private arguments. The real source and relocated Linux native theme/multi-publisher/private-request/ reload/restart scenarios each pass eight cases, with native and selected controls, both PTYs exiting 0 and sixteen retained captures. The same archive passes all 64 earlier native checks. The exact 6ff63f4 candidate passes all four native platforms, with 260 downloaded OpenTUI checks and 32 installed-runtime smoke reports independently audited. The earlier cdb2d4a macOS bench failures remain preserved, with exact ANSI/input diagnoses and scoped acknowledgment repairs. This closes the appearance checkpoint; pure resolution tests are narrower.

Public render adaptation now registers explicit per-pane public slots for conversation, Tools, Help, status and ordinary plugin views. The runtime keeps normal chain ordering and render-only effect guards; the host validates final trees and fences asynchronous results and displayed actions by current source, viewport and generation. Native default-scroll state and resident private input owners remain independent of replacement public content. Two public normal-mode status rows expose the real bundled theme hook. Local Tools diagnostic arguments from UI operations never enter the public render projection. The complete actual CLI scenario passes eight cases from source and a relocated Linux native archive (native five, selected three), including exact action, privacy, coalescing and stale-generation receipts. The same immutable archive passes all 72 preceding native cases. Independent local archive, receipt and screenshot review passes; exact-head four-platform verification remains required before closing this gate.

Protocol 0.7 now has data-only native presentation/input messages and an actual application broker. The production OpenTUI path serializes and decodes those messages through an asynchronous stream; resident widgets no longer call host ports directly. The separate complete grid client retains the compatible 0.6 message semantics. Source CLI navigation nine and native/selected public-view eight scenarios pass through the new production path, including batched command and paste, private fields, reload/resume and restart. Exact-head four-platform verification is still required before closing this checkpoint. The frozen local Linux candidate now passes all 85 checks across thirteen runners, with twenty PTYs exiting 0. This includes five public grid API checks against the renderer module extracted from that exact archive. Full locked coverage passes 2272 tests with three optional skips; typecheck, generation, documentation and the canonical installed-runtime smoke command pass. Independent review of the archive, extracted members, source hashes and screenshots passes. Exact-head remote verification remains required; these local results do not establish physical-terminal or broad 1.0 readiness. The preceding 730d86c release run remains failed: both Linux jobs pass, while the two macOS appearance checks read saved layout before asynchronous persistence settles. Screenshots and final files prove the correct Tools show/hide effects; the corrected bench waits for those exact saved states. It still needs a fresh matrix and does not turn the failed run into success.

The subsequent 3281bda checkpoint locally verifies the production protocol and exact delegated/managed request attribution, including the original selected controller leases. Its exact-head source/platform qualification remains open as recorded above. Physical terminal/media coverage is still required. Ordinary UI operations and retained plugin services now receive independent approval/question/questionnaire/form windows. Verification keeps Main pending while real guarded tools, separate same-owner questions, native field editing and before/after-launcher background input complete. Immediate-command window cancellation now propagates to its exact foreground internal tool or auxiliary model call. Verification includes real worker PIDs, same-owner neighboring commands, held Main input, normal permissions and a reload after operation settlement; explicit background work remains independent. Native terminal/protocol coverage, embedded CPU onboarding and broader real-model self-development quality remain open. These application checks do not establish stable 1.0 readiness.

Historical self-hosting repair diagnostic — 2026-11-13​

Follow-up command-presentation diagnostic: builtin exec and exec_output now send readable stream sections to the model while retaining raw structured results and host-derived exit-status receipts. One captured failed-test result shrinks from 12637 to 7090 bytes; this is a formatting measurement, not a quality or billing result. A fresh expanded-tool Gemma run on the same historical task made 30 requests and changed only plugins/todos/hooks.ts, then stopped at the provider's 4096-token output limit. Independent acceptance still fails six of seven checks; typecheck passes. All 807 baseline files, 21 local workspace aliases, frozen acceptance files and fixture Git HEAD were audited. This is another failed repair, not evidence of self-development readiness. Retained artifacts are under dist/self-host-readable/.

The default strategy now permits one bounded recovery from a provider length finish, discarding all calls in that response and recording a replay-safe context boundary. The retry consumes an ordinary step under unchanged output/token limits. Ten regressions cover effect admission, replay/resume, permissions, budgets, interruption and storage failure. In a fresh historical Gemma diagnostic, recovery occurred once and execution continued, but the task still failed at 60 steps: 78 requests, six compactions, four production files changed, six of seven acceptance checks still failing, and an introduced type error. All 807 baseline files, frozen inputs, 21 workspace aliases and fixture HEAD were audited. Artifacts are under dist/self-host-output-recovery/. This proves recovery was exercised, not task-quality improvement. Next investigate the observed edit-argument, ambiguous-match and stale-snapshot failures and verify complete repairs. Stable 1.0, broader model coverage, onboarding and OpenTUI remain incomplete.

Guarded-edit feedback follow-up: malformed ranges/missing guarded hashes now fail before approval, and match errors distinguish counts/bounds with explicit UTF-16 units. Replaying three saved failures on original source confirms that their oldText matches once when the original read range is retained; operator replay is not a model repair. In a fresh Gemma condition the original seven checks reach 6P/1F with types passing, but the 60-step turn still fails (73 requests). A separate bounded continuation also fails at 60 steps (72 more requests), retaining 6P/1F. The remaining original failure is project-specific Todos persistence. These are partial results, not successful self-development or a causal quality comparison. Both episodes audit all 807 tracked files against allowed changes and preserve original acceptance files, configuration, Git HEAD and 21 local dependency aliases; see dist/self-host-edit-feedback/.

The model's patch also hardcodes some invalid-field indices. A separate ten-case audit of the original exact-field requirement passes only three cases on that patch and all ten on the independent reference. These are separately maintained regressions in packages/app/test/task-plan-validation.test.ts; they do not alter the frozen seven-test diagnostic score. Complete repair must satisfy the actual input-validation requirement, not only those seven examples. Repeated invented edit ranges and incomplete/generalized repairs remain next investigation targets.

Portable snapshot-reference follow-up: builtin read and supported Callix read_symbol results now provide editRef, copying the exact existing guards as one string. Explicit anchors remain supported, references grant no permissions, and malformed/conflicting/stale or wrong-path references are rejected. Session replay, workspace isolation, policy and native Callix tests pass. The full suite has 1475 passing tests plus three optional skips; an explicitly enabled real Callix navigation run passes all 21 cases.

A fresh historical Gemma condition copies issued references in all three edit calls (two applied, one stale rejection, no invented numeric ranges). Nevertheless it fails after 35 requests/27 steps on a second length finish: 6F/1P on the original acceptance, failing types, 0P/10F on the separate exact-field audit. The manual reference passes 7/7 and 10/10 respectively. Audits cover 807 tracked files, frozen inputs/configuration, unchanged Git HEAD and 21 local workspace aliases; artifacts are under dist/self-host-edit-ref/. This is mechanism evidence, not improved repair quality or stable self-development.

These matched diagnostics explicitly cap output at 4096 tokens. The standard TokenBudget reserve at the same 32768-token window is 8192 before provider ceilings or explicit overrides. Next evaluate that separately labeled default reserve condition without raising the cumulative task allowance or rewriting the failed diagnostic's result. Stable 1.0 and the broader remaining scope stay open.

The separately labeled standard-reserve run sends 60 main requests with an 8192 limit and 21 summary requests with a 2048 limit under the same 1.2M cumulative allowance. It reaches 60 steps, with nine compactions and one length recovery: 5P/2F on the original suite, 10P/0F on the separate exact-field audit, but failing types. Four source files change; all 807 baseline files, frozen inputs/config, HEAD and 21 workspace aliases are audited. Artifacts remain under dist/self-host-default-output/. The task is still incomplete; this is not a causal comparison with earlier random model runs. Seven of 20 edit attempts fail because the model omits path while supplying a reference that already encodes it; effective-path derivation before permission review is the next concrete tool issue.

An independent deterministic regression also exposed a completion-policy hole: an edit erased earlier executed-check obligations, so unrelated passing lint could hide an earlier failed test. The policy now retains non-refused attempts within the turn and requires each same check to pass after the latest edit. Pre-edit refusal-only checks retain their prior exception, while mixed executed/ refused obligations still receive Do-not-rerun guidance. Four regressions and the full 1479-test suite pass. This closes that false-completion path; it does not make arbitrary recognized commands a correctness proof or mark the model repair done.

Reference-only target follow-up: edit now derives omitted path from editRef before native pre-tool hooks, rechecks rewritten inputs and supplies canonical paths to native file-policy hooks. Provider arguments/metadata survive replay; the journal separately records the effective target and host projection marker. Real checkpoint/rewind and scope-guard tests pass. Additional regressions reproduce symlink guard evasion and native operations following a changed symlink after approval. Native execution now stays bound to the approved canonical path and rejects a changed resolution before its file service. Full suite: 1494 pass, three optional skips; the explicit native/navigation/core run passes all 36 cases.

Two new historical model conditions still fail without editing: Gemma makes 24 requests then hits bounded empty-completion failure; DeepSeek, using default reasoning and request timeout in a separately labeled condition, makes 11 requests then fails text summarization. Both retain 6F/1P, pass types and fail all ten separate exact-field cases; references pass 7/7 and 10/10. Audits cover all 807 tracked files, frozen/config inputs, HEAD and 21 local aliases. Artifacts are under dist/self-host-ref-target/ and dist/self-host-ref-target-deepseek/. They predate the final approval-path binding correction and do not exercise model edit usage.

Content-free DeepSeek stream statistics show summary retries ending in length or tool_calls, including the final rejected tool-call finish. No summary tool was executed. The default summarizer now requests explicit disabled-tool mode through all four bundled protocols and rejects tool events even with a misleading stop or length finish. Local protocol and compaction-retention tests pass; the full suite passes 1514 tests with three optional skips. A proposed live replay of the saved transcript was rejected by automatic approval review because that context could be relayed through Ollama to a cloud model with unverified handling and billing. No replay request ran. Obtain explicit approval for that payload/egress before replaying it, then investigate provider compliance and bounded reasoning/output handling. These observations do not establish a completed repair or broad model quality.

Compositor was asked to repair the actual Todo scope/status defects in an isolated f1557fb source snapshot, with the two immutable acceptance files from the later manual repair. Workspace dependency aliases point into that snapshot, preventing the current fixed packages from accidentally satisfying its tests. The negative control fails six of seven checks; a separate manually repaired reference passes all seven. This is a known historical task, not an unseen benchmark.

Four live DeepSeek-v4.1-flash runs through loopback Ollama failed without editing source. The first two exhausted a 240000 cumulative-token allowance; default 200-line read pages did not establish a task-quality gain. A separate 1200000-token condition failed when a summary merge did not shrink. After adding one bounded merge correction round, the same larger-limit condition completed five compactions but reached its 60-step limit with 51 reads across 17 paths and no edit. The independent acceptance checks still fail six of seven. These are failures, not self-development success; model variation and the changed allowance do not establish a causal quality improvement.

The default reader now returns 200 lines with a continuation offset, retaining explicit larger ranges and exact edit anchors. Summary correction preserves the original transcript on repeated failure. Both mechanisms have deterministic regressions. Next investigate repeated research after compaction and validate actual source edits and verification on the unchanged task. Evidence is in dist/self-host-task-state/, including all four trajectories and a comparison. No OpenAI request or spending-guard change occurred; Ollama monetary billing is unavailable. Stable 1.0 and sustained self-hosted development remain open.

A follow-up adds bounded host tool observations that survive independently of the natural-language summary and stay hidden from the chat UI. Deterministic checks cover false summaries, repeated compaction, replay/resume, output bounds, failed calls and history restoration. A live DeepSeek run retained one record through each of six compactions, but still reached 60 steps without edits. Gemma made four-file changes in a separate 60-step run, with five acceptance failures and failing types. A separately bounded continuation through saved-session resume also failed those checks, including repeated rejected no-op edits. Neither episode is a completed repair or a quality gain attributed to the new record. Tests/configuration and dependency isolation remained intact, and no model-authored patch was applied to the main checkout. Evidence is retained in dist/self-host-task-state/evidence-receipts-comparison.json; both model families and the failed initial/continuation results remain visible. Next investigate repeated failed edits and preservation of useful implementation progress.

One later Gemma continuation uses early argument admission: four identical exact edits are rejected before tool.check or permission review, preserving whitespace normalization for unguarded calls. It still reaches 60 steps with five acceptance failures and failing types. Its six allowed source changes remain isolated; tests, configuration, Git HEAD and workspace dependency aliases are intact. This is a separate continuation from a later partial patch, not a paired comparison or a successful repair. Evidence is in dist/self-host-task-state/attempt-edit-validation-gemma-resume/. The new validator improves admission semantics; broad live model repair quality remains unproven.

A separate expanded coding diagnostic uses a new clean snapshot, 14 single-agent tools, automatic project instructions/skills/references, native Callix navigation and project_check. The entire evaluator process has a minimal environment, including SDK-owned subprocesses. Zero-model preflight discovers 59 npm scripts, runs real typecheck and locates ToolContext through the native semantic index. The original six-failure baseline and separate seven-pass reference remain intact. One live attempt incorrectly restricted edits to package sources while excluding Todos plugin code; its four refusals are retained as a test-harness scope error. A separate run explicitly authorizes Todos production TypeScript and excludes test/spec files. It uses semantic navigation once and changes two source files, but exhausts its token allowance with the same six acceptance failures and failing types. Neither run uses project_check autonomously. More available tools alone do not establish a quality gain. Frozen tests/configuration and all 21 workspace aliases remain intact; evidence is in dist/self-host-expanded/.

This investigation also exposed a real extension composition defect: structurally valid class-based tools lost prototype methods when copied. A disk-plugin regression reproduces that failure; core now retains class/private receivers, non-enumerable methods and metadata getters while preserving native function identity and ordinary object behavior. The loader also types optional validators and rejects non-function declarations without blocking legacy peers. Real load/reload/admission tests cover these contracts. These are extension fixes, not successful model-authored repairs or stable 1.0 completion.

Offline CPU coding diagnostic — 2026-11-13​

The maintained local CPU evaluator verifies an existing pinned Qwen cache, owns its loopback runtime and runs the actual app with an explicit focused tool composition. Preflight rejects absent behavior and accepts the scripted reference using the unchanged checkout-case grader. Four working-tree configurations were observed twice: general instructions, missing-edit feedback, coding instructions, and coding completion feedback. All scored 0/2 on this known task. Files covered by preservation assertions survived; runtime records confirmed CPU and cleanup. No paid model call was used. These are development diagnostics, not held-out quality gains or evidence of Claude Code parity.

The coding-instruction runs exposed an unchecked completion path: without a saved plan, a model could write an unrelated file and claim checks passed without executing them. A manually authored harness repair now requests a successful recognized check after the latest built-in write/edit for the coding preset. Started background jobs do not count until matching terminal output, and passed checks before a later edit are stale. Deterministic regressions cover opt-out, general behavior, failed/unfinished/killed jobs and project_check cooperation. The live post-repair attempts still failed and did not perform a successful edit; they do not establish live repair success or directly exercise that write gate.

An independent installed-binary probe rejects the previous archive's unchecked completion, then verifies the new archive's bounded failure and a real-check success while preserving the user draft and checker. It is now part of every native release smoke. This demonstrates completion-policy behavior, not autonomous model coding ability. See offline CPU diagnostics and coding completion checks.

On the local Qwen 3 1.7B checkout case at 1c1c7b5, four identical-text edits failed, neither project checks nor other writes ran, and the model claimed a successful fix. The independent grader still got 18 for an expected 12 after nine model requests. The coding completion policy now requests a correction or an explicit unchanged outcome when all built-in mutation attempts failed. App regressions exercise repeated false completion, a reported blocker and recovery through a changed file and a successful check.

A second live attempt used the updated policy, clarified edit instructions and a fixture containing only its original four files. It spent all 12 admitted requests on discovery. Eight grep calls treated @. as a literal working directory. The grader again got 18. This attempt never proposed completion. Callix now recognizes that exact working-directory argument when the primary project has no literal entry of that name. App tests cover the primary and attached projects, a real directory with that name, and invalid symlink targets. In the final live run, all four searches used the default directory and four reads reached the two implementation files. All six edit requests omitted both path and editRef. The model exhausted 12 requests without a successful edit or a project check. The independent result remained 18. The runs used different host code and do not measure a quality improvement.

The six targetless calls in the last checkout attempt contained only oldText and newText. Those were the editor's only top-level required fields. Its path-or-reference requirement depended on a root JSON Schema anyOf. The pinned Qwen chat template renders the Ollama function structure, and the locally inspected Ollama root parameter representation has no field for that alternative. The native editor now advertises a required string target alongside both text fields. Each local CPU transport receipt records those required fields for comparison with the tool calls the model produced.

In a fresh local checkout attempt all 12 HTTP requests advertised target, oldText and newText as required. Qwen selected math.mjs and changed addition to subtraction. The independent behavior check passed, and the protected package file and tests retained their original bytes. Five subsequent edits repeated the replacement and found no matching text. The model exhausted its request allowance without executing the project checks, so the overall case failed. The core now includes the path of an earlier successful identical replacement in that specific no-match error when the corresponding tool result remains in the turn. This result does not prove the file is still current or that any tests passed. A separate local greeting case on the required-selector schema also used all 12 admissions. It made nine searches and three reads, left the source unchanged, and failed both the original checks and the independent behavior grader. Neither episode qualifies as a complete autonomous task.

A further CPU checkout run on clean c458dc7 repaired the arithmetic and then used write with just the existing checkout function. That discarded its import. The agent ran npm test twice and both executions failed on the missing discount binding. The native writer now rejects a verbatim excerpt submitted as a whole-file overwrite unless intentional extraction is explicit. A regression replays the operator fix and the erroneous write, then runs the real project check and the original independent grader.

The next live run took a different path. It used edit to introduce a duplicate checkout declaration, duplicated the incorrect discount function, and ran npm test twice. Both attempts failed on the duplicate declaration. The 12-request limit was reached and the independent behavior check failed. This live run did not exercise the excerpt guard. Both runs preserved the protected tests and package file. See the current write contract.

Acceptance gates​

On 2026-11-18, an installed Linux x64 Compositor authored the known initial Project checks task through the saved OpenAI gpt-6.1-sol profile. Ten model requests settled successfully after fixing Responses zero-hosted-tool counters and cache-write lifetime metadata validation. Protected project files survived and both required project commands passed, but independent loading rejected the candidate's nonexistent ui.open hook. An extra agent-selected test command was denied by the declared permissions. This is a failed authoring result, not a completed useful-plugin delivery. Report and candidate bytes are retained in dist/development-openai-live-20261118/. The user authorized a $60 total ceiling; the original ledger and pending reservations remain intact.

GateRequired evidenceCurrent assessment
Real developmentRepeated end-to-end bug fixes, feature work, refactoring and plugin development on realistic repositories; independent behavior checks, preserved user changes and successful project verificationLive Compositor-authored fixes, a guided plugin review and a multi-file form refactor have independent artifact checks. The form episode also exposed a harness typing defect and reported it truthfully. Sustained self-hosted development and broader task coverage remain unverified.
Model independenceContract tests for every supported protocol, plus repeated real-model development runs across different model families. Unknown capabilities, invalid tool calls and weak tool use must degrade with bounded recoveryHTTP fixtures cover the supported transports. The 0.8.0 known-task diagnostic scores 6/8 across Gemma and DeepSeek, with both families scoring 3/4. Saved-profile OpenAI evaluations are available. Broad live model quality remains incomplete.
Tool cooperationFile discovery, semantic/text navigation, precise edits, test execution, shell jobs and delegation reach verified artifacts without unnecessary retries or denied-action bypassIndividual tools and regression suites exist. The 0.8.0 two-family diagnostic exposes repeated denied scratch-test execution and a turn token allowance refusal before completion. End-to-end efficiency and recovery need wider evidence.
Long-running executionContext pressure/compaction, interruption, steering, queued work, restart/resume and failures at effect boundaries preserve the task, scope and truthful completion stateOrdinary queued prompts remain in memory while waiting: interruption cancels them and accepted user messages become reviewable immediately; new journal records correlate new, queued and steering user inputs with committed submissions or applied steering by ID and /recover offers unmatched entries for explicit review, retry or discard; entries remain ambiguous when cancellation or interrupted hooks were not recorded, and older entries cannot be identified. Broader process-death and multi-hour queue recovery remain required. One live known task passed process-death/resume/explicit-compaction checks. A second known task recovered from a failed summary and passed three automatic compactions in a configured 8192-token window with independent artifact checks. A controlled live API truncation probe also passed. Multi-hour development and broader repositories remain required.
Extension qualityA developer creates, installs, enables, configures, updates and removes a useful composite plugin through Compositor. Public APIs remain compatible; disposal/reload and custom storage work in native archivesGit review passed source and exact archive certification on all four native platforms at c98b3ce. Project checks 0.1.1 passed the same platform scope at f102fef after fixing macOS project-path aliases. A completed autonomous initial plugin delivery and broader extension coverage remain required.
User workflowThe real terminal supports keyboard/mouse navigation, concurrent plugin/settings work and pending approvals/questions, with cancellation and narrow-window behavior verifiedReal PTY scenarios cover concurrent management and a complete 48×18 question/management/answer flow with draft preservation. Broader daily-use and terminal coverage remain incomplete.
Trust and resource controlPermission, scope, secret handling and source-injection scenarios preserve boundaries. Model calls, retries, workers and shell processes have enforceable budgets and cleanupExisting guards and tests provide evidence for specific boundaries. Permission rules are not an OS sandbox. Threat-model review and broader adversarial sessions remain open.
Release qualityAll required tests, documentation, plugin certification, four-platform native builds and installed-archive smoke pass for the exact candidate. No known critical workflow failure remainsv0.8.0, revision 4b5b297, passes candidate and tagged four-platform build/install checks, all 32 tagged runtime reports and checksum verification of all six public assets. Public Linux 0.7.4→0.8.0 upgrade preserves settings, prior binary and session context. Setup, independent conversation tabs and workflow PTY evidence matches public shipping files. Broader 1.0 gates remain open.

At source revision a243d9c, the actual PTY conversation-copy scenario passed both native and selected-editor compositions (eight checks total). It covered reopening closed windows, shared input and cursor, pending private requests, reload and session history after restart. The reports and captures are in dist/goal-opentui-copies-current/. This does not establish persistence of an unsent Main draft across process restart or certify a native release archive.

Each gate needs links to source, commands, immutable revisions and retained run artifacts. A narrow passing check cannot close a broader gate. Missing, contradictory or non-reproducible evidence leaves the gate open.

The built-in OpenTUI Main editor now persists unsent text per session. Ink and selected plugin editors still keep their ordinary Main drafts in memory. Durable recovery must distinguish a draft from identical input already admitted to the session, including a crash between journal admission and editor acknowledgement. The identical-input regression queues two identical messages, retries one by its recovery ID and checks that the other remains available after journal replay. Matching by text alone would conflate them. Core can now reserve a single-use input identity before starting admission, and controlled tests correlate it with new, queued and steering intent records, including a late write after interruption. The built-in OpenTUI editor uses that capability; the Core tests alone do not establish all crash boundaries. The private storage component has focused tests for per-session files, bounded input, malformed-file preservation and correlating a prepared ID with the exact journal intent and text. Each instance owns its versions; divergent versions are returned for explicit review, and resolution requires the exact versions reviewed. Tests cover two separate local processes and a version changing during review. A separate test kills a local writer after its save and reads and edits the draft from a new process context. These checks do not establish network-filesystem or arbitrary crash durability. These component tests alone do not establish the visible recovery workflow. The replacement storage format writes immutable version files. Saving automatically compacts after 64 versions; explicit compaction is also available. Both remove only versions superseded in a validated snapshot, while preserving conflict leaves; tests cover concurrent writers, clearing and malformed files, 150 edits with automatic compaction, a separate compactor process while a writer saves, and process death immediately after either of two deletions. Separate process-death tests cover the atomic publication boundary before and after rename. On Linux, explicit compaction can remove temporary files from a confirmed dead process in the same boot and PID namespace. A test pauses a live writer, then kills it, and verifies that only its attributable file is removed. Legacy, foreign-domain and unidentifiable temporary files remain; other platforms do not perform this cleanup. This is logical file deletion, not secure erasure from storage, backups or filesystem snapshots. The old experimental v2 files remain readable and are never deleted by v3 compaction. A continued legacy writer produces a conflict when its version changes; tests cover that case, multi-generation compaction and malformed legacy data. Compaction also preserves an intermediate record when it is the only link suppressing a retained legacy ancestor; a focused regression checks that a later edit permits compaction without resurrecting the old text. Old files can retain private text and still need a retirement policy. Broader interruption, filesystem and retention-policy coverage remain open. A session-scoped coordinator now saves a prepared input identity before calling Core admission, blocks stale revisions and retains ambiguous input for explicit review if admission or clearing fails. Focused tests cover queued admission, clearing failure, session replacement and conflict refresh without calling a model. OpenTUI's host now binds the coordinator to each Core generation and accepts negotiated, revision-checked changes for the built-in editor. Once a client opts in, normal prompts and host commands are armed before dispatch. Host and wire tests cover a resumed session, stale ports, a prompt and a command. The built-in renderer now consumes draft state, serializes edits around Enter, and presents explicit review and conflict selection. The PTY scenario verifies normal restart, SIGKILL after a saved edit, ambiguous input, fast Enter with later typing and conflict selection with a loopback provider. These are scoped checks, not arbitrary crash or power-loss durability; selected editors and broader terminal coverage remain open. The native presentation protocol negotiates draft exchange and a separate optional review capability. The latter carries conflict variants and explicit review, resolution and refresh actions; older peers receive only the fields they understand. Host, client and wire tests cover stale versions, hidden windows, conflict refresh and explicit review without a model call. The built-in renderer uses these operations. Nine dedicated PTY scenarios passed against a locally built, relocated Linux x64 executable (SHA-256 34ce8da1eace32fc9e16861e1abbc91e101d626748ad147d073e09d6ecde5ca2): normal resume, review by keyboard and mouse, conflict selection, a second conflict page, conflict selection by mouse, SIGKILL after a saved edit, Enter followed by later input, and a normal Ctrl+C exit with a pending edit. The external renderer's SHA-256 is cd7977d88bbf53d2e94891804856124c5f2af563887dc9586b9511a871ded196. Reports are in dist/goal-main-draft-ultimate-*. They do not certify other platforms or a physical terminal. The scenarios are included in source CI and the four-platform installed-archive workflow for future runs.

The check/report cooperation test runs actual Vitest through project_check, reads its generated JSON through test_report, repairs the controlled fixture and repeats both tools. It checks that a failing check remains a failure, the second report reflects the repaired tests and an unrelated user file survives. This is a deterministic tool pipeline with zero model requests; the test host supplies the repair. It does not establish autonomous diagnosis or close the broader tool-cooperation gate.

Evaluation rules​

  • Judge actual artifacts and independent behavior, not the agent's completion text or a success marker printed by an arbitrary command.
  • Freeze tests only when the task requires that. Otherwise permit useful test additions while independently executing the original checks and extra probes.
  • Keep grading sources and credentials outside the model workspace. Clear host credential, proxy and preload values from project-check processes; this is environment hygiene, not filesystem or OS containment.
  • Record model, provider protocol, budgets, request counts, token usage, failures and repetitions. Separate deterministic integration scripts from autonomous live-model runs and development cases from unseen evaluations.
  • Preserve failed trajectories and explain rubric corrections or source changes. Do not call a run an unbiased improvement comparison when the task, rubric, instructions, model or implementation changed between measurements.
  • A comparison with another agent needs comparable tasks, model access, budgets, permissions and grading. Until those are measured, do not claim Claude Code parity or universal quality across arbitrary models.

See evaluation commands and limitations, extension certification, execution contracts and release gates.

Current development focus​

The initial test-reports backend tool reads real local Vitest/Jest JSON, derives assertion totals and returns bounded failure records through normal canonical-path permissions. Its independent package checks include explicit activation/removal, malformed input, UTF-8 byte bounds, symlink denial and cancellation. The source package now also provides a manually implemented /test-report panel with explicit reads, failure selection, scoped defaults, restart and concurrent-agent/draft-preservation checks in the actual PTY. The composite lifecycle exercises the tool, command, view and configuration across install/reload/restart/update/rollback/disable/re-enable/remove. The maintained terminal case now requires [x] and Alt+Q to settle the owned operation while preserving the main model/draft, and requires reaching the final diagnostic line and returning to the path in a 60×24 split. The host keeps late publications hidden until an explicit plugin command/action. Opt-in host paging is an additive catalog0.24 field; legacy input routing stays unchanged. These checks do not establish autonomous plugin authoring. The required host repairs ship in 0.7.4. The independent report-plugin 0.1.1 release at 4358ed1 passed source and all four native archive certifications and tagged publication. Five public certificates, 141 evidence references and the public Linux archive passed independent checks, including 13 actual report/package behavior cases without a model request. A fresh live Compositor authoring attempt used a minimal task-specific profile with ordinary read/write/check tools and no reference implementation in its workspace. Three admitted requests only read repository context; the fourth request exceeded its predeclared serialized-body envelope before admission. No candidate files were written, all protected files survived and the package oracle correctly failed. The attempt does not establish autonomous initial plugin delivery. Do not infer model quality from this bounded attempt, expand its budget retrospectively, or count the manually implemented package as agent-authored. Full plugin/UI authoring, sustained development and the wider model/task matrix remain open.

A bounded Sol editing turn through Compositor added visible CI-panel busy feedback using one guarded edit and two model requests. The host supplied a real read snapshot; the agent had only the scoped edit tool and correctly stated that it did not run tests. Three frozen success/failure/cancellation checks failed on the original code and passed independently after the edit. All 735 protected files and the oracle survived unchanged. The real PTY flow also verifies visible feedback during a subprocess, status removal after cancellation, preserved old results, concurrent main-agent work and restart behavior. This is one small known UI development case with supplied context, not autonomous investigation, initial plugin delivery or proof that the broader usability gates are closed. The busy feedback ships in the independent CI-plugin 0.1.1 release; the immutable 0.1.0 archive remains unchanged.

Preparing that first update exposed a certification gap: native archives could omit previous-release transitions even though source release checks required them. Artifact certificates now also require upgrade/rollback evidence against their platform's pinned previous archive. A regression rejects missing scenarios, wrong versions/hashes and reused evidence from another platform. The CI scenario uses real 0.1.0 release bytes, normal marketplace install/enable/update/rollback, two restarts, retained settings/draft and actual tool/command probes without model requests. Revision 9a2314e passed source and all four native candidate gates, Node 22/26 CI and documentation checks. Its tagged publication workflow also passed. Five public certificates, 136 evidence references and four archive checksums were independently checked. Public Linux bytes passed the actual manager's install/update/restart/rollback/restart flow through a local verified-byte mirror, preserving settings and draft without downloading during rollback. Visual review caught stale screenshot capture and early cancellation acknowledgment in the bench; both now synchronize on the actual rendered state.

Initial CI-plugin authoring through Compositor exposed two evaluation-fixture settings mistakes, then a runtime compaction gap. The latest user request was retained for the main model but omitted from auxiliary summaries; scaffold examples were repeatedly mistaken for the actual task and its approval state. The source repair passes the exact request separately to every summary segment and merge, with context accounting and fail-closed handling if it cannot fit. Deterministic regressions and one controlled live summary probe pass. A full development rerun from an empty candidate under the repaired runtime was subsequently attempted at fbd69d6. It stopped before request admission at the authorized USD guard after six main steps, three compactions and thirteen reported requests, without writing candidate files; all 726 protected files and both frozen oracles survived. This budget-curtailed attempt does not establish initial delivery. A guided completion from the saved CI candidate finished in seven model steps, passed the frozen ten-case oracle and strict compilation, and preserved all 725 protected files. This is a continuation of existing code, not a successful autonomous initial delivery. Independent review then repaired stale repository results, mismatched run IDs, incomplete bounded JSON records and the UI hook timeout. The integrated /ci candidate passes source PR certification, lifecycle/update/rollback checks and the actual PTY concurrent configuration/run/job/log/cancellation/restart flow. Two actual read-only GitHub CLI requests through the saved proxy also pass with zero model requests. The exact fbd69d6 candidate passed source and all four native installed-archive certification jobs; five certificates and 131 evidence files were independently verified. The subsequent tagged run failed a PTY synchronization assertion, so publication was skipped. The scenario now waits for the specific refresh response to render before issuing another request; its previous process-start/old-result predicate could accept stale UI. The corrected real PTY flow and full local source certification pass at 8757290. The corrected 14655ca candidate passed source and all four native archives, with five certificates and 131 evidence files independently verified. The failed unpublished tag was moved to that certified revision after checking release absence; the subsequent tagged verification and publication passed. CI review 0.1.0 is published independently. A public Linux x64 download through the saved proxy passed checksum verification, inactive installation, explicit activation, /ci without a model request and disable. This supplements native archive checks; it does not establish public installation on every user's machine. This repair is not part of immutable v0.7.1.

Further actual-core regressions exposed task-reference replacement by host completion/recovery feedback during compaction. An explicit harness-origin marker now keeps those messages as working context while preserving the actual user task for summaries and permission review; real steering still changes the task. This includes literal user text beginning with the harness's wording. One oversized message can be folded when it is the only remaining middle entry, subject to the existing reduction/failure guards. Automatic-compaction, replay and permission-context regressions pass. The origin repair ships in 0.7.2. The maintained installed-binary regression was added later at 026e605 and passed a separate four-platform candidate.

A later source fix extends task preservation to plugin-owned automatic follow-ups. Controlled regressions first reproduced an automatic build notification replacing the actual task for permission review and compaction. Queue ownership now marks that message as host feedback, with journal provenance preserved on replay; initial delegated tasks and literal user steering still count as requests. Session request previews also exclude host feedback. The repair ships in v0.7.3. Its tagged native matrix passes an installed-binary background/replay scenario on all four platforms, with the actual task retained in two permission reviews and every auxiliary summary. A separate startup witness checks immediate EOF without a model request and initialization before the first plain-session model request. Public Linux bytes pass the compaction/background/replay probes after an installer upgrade from 0.7.2 with state and the old binary preserved. These are controlled local transport cases, not broad long-session or adversarial-model coverage.

Local Qwen3 1.7B CPU diagnostics at 0b48eca used three fixed small tasks with two repeats each. Default thinking passed 4/6 strict cases; explicit reasoning_effort: none passed 2/6, both with minimal instructions and with the stock agent harness. All protected fixtures survived. Preemptive refusal of the known forbidden write was safe but did not exercise the rubric's actual host denial requirement. The edit cases exposed output-limit stops, an unnecessary rewrite that lost a trailing newline, and empty successful-looking final turns. These are controlled small-model diagnostics, not autonomous development or a broad model-family quality result.

A fresh Sol bug-fix run through Compositor then repaired empty final responses in the replaceable finish policy: the existing two bounded continuations ask for a useful answer, then fail if the model remains silent. Disabled verification and replacement policies remain supported. Six model steps/requests completed with the frozen five-case acceptance suite and typecheck passing independently; all 732 protected files survived. The integrated change passes 1106 maintained tests with three optional skips. A live replay with the unchanged stock-harness cases and grader at ca84b9e passed 3/6, with zero empty successful turns versus two before the repair. One edit passed exactly; another correctly reported failed guarded-range edits. This fixes silent feedback, not general editing quality. Remote release certification remains separate; this is another known development case, not 1.0 completion.

At c31c430 the unchanged local cases passed 4/6 with 19 requests and preserved fixtures. Both precise edits still failed through redundant writes dropping a newline. The actionable missing-hash diagnostic was not triggered in this replay, so these results do not establish that it improves live model recovery.

Git review revision c98b3ce passed full source and four native installed-archive certification. The downloaded certificates, evidence hashes and embedded archive identities were independently checked; native macOS keyboard/focus and cancellation screenshots were inspected. This manual candidate run did not publish a plugin release. Project checks' concurrent candidate exposed a real macOS path-alias rejection; the selected project root and package path now both use host canonical resolution. An aliased-directory integration test also rejects an outside package link even when its directory is attached to the host. The corrected Project checks 0.1.1 revision f102fef passed source and all four native archive jobs. Independent checks validated its exact source/package identity, evidence hashes, archive checksums and embedded release metadata. macOS ARM64 success/narrow and Intel cancellation screenshots were inspected. These are deterministic/tool and real-terminal cases; they do not establish autonomous model development quality. The candidate run skipped publication and does not certify later source revisions.

A subsequent multi-file form refactor used project_check to discover and run real project verification scripts. The model preserved the frozen acceptance file and controlled user draft, passed 68 behavioral checks and explicitly reported a compiler failure caused by the acceptance fixture's unsupported String.isWellFormed type. An equivalent fixture assertion restored compiler compatibility; independent checks passed without changing the authored source or making a paid rerun. This is development and truthful-error-reporting evidence, not an all-green autonomous episode under the original defective harness.

The published v0.7.2 patch is bound to immutable revision 888bf8c. Its Node 22/26 CI, documentation checks, four-platform candidate and tagged publication passed. Two macOS Intel attempts failed on DNS: first during artifact upload after successful smoke, then during the Callix download before compilation. Those failures were retained; a rerun of the failed job completed all required checks and publication. No failed functional assertion was hidden by retries. An independent public Linux update from 0.7.1 verified installer/archive checksums, preserved the previous binary and application state, and passed the context regression on the publicly installed executable. This is a release checkpoint, not completion of the broader development/model/long-session gates.

The earlier published v0.7.1 patch is bound to immutable revision 7e8e389. Its Node 22/26 CI, documentation checks, build-only four-platform candidate and tagged four-platform verification and publication passed. This patch fixes interrupted-turn settlement/cancellation, accepted background follow-ups, project-root aliases and CI terminal interaction, and ships bounded installer downloads. These are verified runtime/distribution fixes; they do not close broader model, self-development or long-running gates. An independent Linux x64 check downloaded the published installer and assets, verified checksums and upgraded an isolated 0.7.0 installation to 0.7.1 through the saved tool proxy. The old binary and application-state sentinel survived. This public-download check supplements the four native CI install checks; it does not establish public installation on every user's machine.

The earlier published v0.7.0 milestone points to immutable revision 21d7934. Its Node 22/26 CI, documentation checks and release verification, four native build/install jobs and publication all passed. The release contains four platform archives, the installer and checksums. Independent plugin archive publication and the broader 1.0 gates remain separate.

The previous immutable candidate a9f99c9 has green CI and four-platform candidate build/install smoke. That build-only workflow did not publish a new application release.

The first work toward this goal strengthens the evaluation foundation: explicit preserved fixtures, independently rerun original checks, correlated verification results, sanitized grader/shell environments and saved OpenAI profiles with a persistent request-admission budget. These make subsequent quality work measurable. They do not close the 1.0 acceptance gates.

Next, expand beyond small fixes to multi-file feature work and refactoring, then exercise interruption/resume, context pressure, parallel agents and plugin creation as complete development workflows. Use failures in those runs to drive runtime and tool changes, and verify those changes against unchanged criteria.

A second authoring episode produced the Project checks plugin through Compositor. Its unchanged six-test acceptance suite and five agent-added regressions passed, but the turn ended at the conservative USD guard before a final response. Independent review found and corrected metadata bounds, settings scope leakage, command usage, cancellation placement and hook-expiry cleanup. The original candidate and traces remain separate from the reviewed integration. A later live guided review iteration from the original candidate completed in 22 model steps with two automatic compactions and passed all 18 checks plus typecheck on independent rerun. Its reviewed settlement logic is integrated. This is known-task review evidence; the completed autonomous initial authoring gate and broader task/model coverage remain open.

Git review adds a concrete tool/panel/configuration workflow and exposed a host subprocess cancellation leak, now covered by real process-group tests. The next runtime work adds concurrent guarded UI tool execution with separate operation state, cancellation, permissions and journal correlation. Source tests and a real Git-panel PTY scenario verify that refreshing and cancelling a real subprocess leave the held agent active. This evidence does not cover every plugin or a live model. Explicit UI command routes now open new panels during a turn; other commands retain their queued prompt behavior. A live known checkout task now passes a process-death/resume/explicit-compaction scenario with independent grading. A later ledger task exposed an oversized-tail recovery gap and a non-shrinking merge failure. The revised compactor recovered the saved task through three automatic compactions at a configured 8192-token window. Frozen files and independent validation, deduplication, sorting and nonmutation checks passed. A separate live API probe forced the first auxiliary output cap to 32 tokens and verified an identical-segment retry ending successfully. These are bounded development/fault-injection cases, not natural million-token or multi-hour sessions. Broader sessions and remaining-platform certification stay open.