Skip to content

CHANGELOG.md

Latest commit

 

History

History
888 lines (618 loc) · 129 KB

CHANGELOG.md

File metadata and controls

888 lines (618 loc) · 129 KB
Raw

Changelog

Flat, reverse-chronological log for cross-AI collaboration (Claude / Codex / Gemini / etc.). Format: ## YYYY-MM-DD — author — headline followed by bullets. Separate from docs/release.md (release-focused) and docs/archive/session-memory-*.md (single handoff snapshot).

2026-07-15 — Codex — docs: multilingual v0.6.31 release closeout

  • Added complete Simplified Chinese and Japanese v0.6.31 release notes, connected all three language editions, and changed each root README to lead readers to the release notes in its own language.
  • Reconciled the current-release references in docs/release.md and docs/roadmap.md so the English, Simplified Chinese, and Japanese evidence surfaces are discoverable from the canonical release and open-work documents.
  • Preserved the release boundary: the immutable v0.6.31 tag still points at the published release commit; these post-publication documentation translations live on main and are linked from the GitHub release body.

2026-07-15 — Codex — release: v0.6.31 published

  • Published and registry-verified @bbingz/polycli@0.6.31 (latest, registry time 2026-07-15T14:37:53.706Z, shasum 57d0f77811767c4310623af03f27af82375abae8), @bbingz/polycli-utils@1.0.5 (2026-07-15T14:37:22.557Z, 99df508a6bffe601e79569927bedf4016d3d471f), and @bbingz/polycli-opencode@0.6.31 (2026-07-15T14:38:07.324Z, 65c990f89df099bb0a1a95104a0a8400abb0f6ca). @bbingz/polycli-timing@1.0.2 was unchanged and not republished.
  • Created lightweight tag v0.6.31 at a70eb093bc7892e2f6b653ed29ca8bba5d66489b and published https://github.com/bbingz/polycli/releases/tag/v0.6.31 (publishedAt 2026-07-15T14:39:17Z; non-draft, non-prerelease).
  • Pre-publication npm run release:check passed (906/906 tests plus strict fixture freshness, source-derived bundle/manifest/host/Codex/review-drift/Claude-plugin checks and npm dry-runs); PR CI run 29421881033 and main CI run 29422037687 passed. Clean registry installation exercised terminal offline agent-context --json (schema 1, build 0.6.31, 20 commands, utils 1.0.5) and imported the OpenCode PolycliPlugin package. Native Windows execution remains unverified and is disclosed in the release notes.

2026-07-15 — Codex — release: v0.6.31 candidate (review remediation)

  • Closed all 14 confirmed findings from the v0.6.30 comprehensive review: truthful no-diff/background JSON v2 results, strict provider target disambiguation, accurate TUI effects, active-job visibility, ledger preview redaction, safe prompt transport, structured session identity, and canonical typed failures.
  • Hardened provider process lifecycle with bounded stdout/stderr capture, POSIX process-group termination and escalation, simulated deadline-aware Windows taskkill coverage, and exactly-once settlement even when close is missing or termination fails. Native Windows execution was not available; streaming paths retain a direct-child fallback there.
  • Made background cancellation and SessionEnd recoverable: cancellation intent remains active until a verified stop, start failures use a private recovery sidecar, terminal ledger/state publication is ordered, and worker/cancel races cannot publish late provider material.
  • Added a source-derived, read-only generated-artifact freshness gate that runs before any in-place bundle rebuild in CI and release:check; rebuilt all five companion bundles from current source.
  • Prepared host/OpenCode/terminal 0.6.31 and @bbingz/polycli-utils 1.0.5; @bbingz/polycli-timing remains 1.0.2 and @bbingz/polycli-runtime remains private.

2026-07-15 — Codex — release: v0.6.30 published

  • Published and registry-verified @bbingz/polycli@0.6.30 (latest, registry time 2026-07-15T08:55:13.583Z, shasum 882e134363d70545c15e060a8da6c1274a2aa1e7), @bbingz/polycli-utils@1.0.4 (2026-07-15T08:51:16.560Z, f89c94947199f4d9d61ec6eddca889bb83a95ec4), and @bbingz/polycli-opencode@0.6.30 (2026-07-15T08:55:44.359Z, 9f71767156d2278f3f9bbe0cadc3fd1c90ae289f). @bbingz/polycli-timing@1.0.2 was unchanged and not republished.
  • Created annotated tag v0.6.30 at c7e6a278542e9761f55c964ef15236417ed81a25 and published https://github.com/bbingz/polycli/releases/tag/v0.6.30 (publishedAt 2026-07-15T09:02:11Z; non-draft, non-prerelease).
  • Pre-publication npm run release:check passed (771/771 tests plus strict fixture freshness, bundle/manifest/host/Codex/review-drift/Claude-plugin checks and npm dry-runs); tag-target CI run 29402849982 passed; post-publication registry hashes matched dry-run artifacts, and clean registry installs exercised terminal agent-context plus the OpenCode PolycliPlugin. See docs/release-notes-v0.6.30.md.

2026-07-15 — Codex — release: v0.6.30 candidate (agent-native CLI control plane)

  • Added one declarative command registry as the source for strict parsing, generated help, host-map validation, terminal metadata, typed errors/output schemas, and the offline agent-context --json discovery contract. Unknown option-looking tokens on registered commands now fail with bounded suggestions; pass -- before prompt text that intentionally begins with an option-like token.
  • Split host, provider, invocation, and attempt identities in state and ledger records; made foreground/background/health terminal pairs recoverable and attempt-correct; decoder overflow now terminates the provider process tree before settling once.
  • Added opt-in --json-v2, explicit --job id:...|prefix:...|latest* selectors with typed waits, and the redacted cursor-based debug tail surface. Existing --json payloads and compatible positional job references remain unchanged.
  • Prepared host/OpenCode/terminal 0.6.30 and @bbingz/polycli-utils 1.0.4; publication evidence is recorded in the subsequent release closeout.

2026-07-15 — Codex — release: v0.6.29 published

  • Published and registry-verified all four public packages as latest: @bbingz/polycli@0.6.29 (registry time 2026-07-15T02:09:46.979Z, shasum c63a5135d77417da46e0b16ef9592d4e74ca5e5b), @bbingz/polycli-utils@1.0.3 (2026-07-15T02:10:38.964Z, 94791ca68cb00f1740f5af540da0f8e29541cb5c), @bbingz/polycli-timing@1.0.2 (2026-07-15T02:10:59.440Z, bd305c872ecd50e0abef6b6bd4abefcc1240e15a), and @bbingz/polycli-opencode@0.6.29 (2026-07-15T02:11:18.658Z, cdf927ecc557602800e18b5feb5f0d3d2e88c0bb).
  • Created annotated tag v0.6.29 at 8f9603480c036b910bc9942195a897037006a6f8 and published https://github.com/bbingz/polycli/releases/tag/v0.6.29 (publishedAt 2026-07-15T02:12:20Z).
  • Pre-publication npm run release:check passed (618/618 tests, strict fixture freshness, bundle/manifest/host/Codex/review-drift/Claude-plugin checks, and all npm dry-runs); post-publication registry tarball hashes matched those dry-runs. See docs/release-notes-v0.6.29.md.

2026-07-15 — Codex — release: v0.6.29 candidate (provider and lifecycle hardening)

  • Prepared the full current release batch: durable background-job terminal transactions and session-artifact handling, status-only default setup auth probes, current provider invocation/parser contracts, and review-safety drift checks. The release keeps the runtime flat and @bbingz/polycli-runtime private.
  • Refreshed parser fixtures from authorized local CLI captures, pinned Claude at 2.1.210, added OpenCode2 preview fixture coverage, marked the Gemini individual sign-in capture route retired, and marked the temporarily unavailable Copilot subscription capture route archived without removing its runtime provider or host plugin.
  • Bumped host/OpenCode/terminal packages to 0.6.29, @bbingz/polycli-utils to 1.0.3, and @bbingz/polycli-timing to 1.0.2; the release gate and registry verification are recorded in the subsequent published closeout.

2026-06-26 — Codex — release: v0.6.28 published (provider-state review)

  • Published v0.6.28 to npm: @bbingz/polycli@0.6.28 (shasum 6ab1bc8e4f8f241ac529058ef325c41f80983e6f) and @bbingz/polycli-opencode@0.6.28 (shasum 0cb4db0ee36d9463122533bf1906ceea954222df), both latest. GitHub release v0.6.28 (publishedAt 2026-06-26T15:21:35Z) + tag v0.6.28. Utility packages unchanged.
  • Closed PR #15 after review and release-prep: main CI green on 5f2cdde8cae50bd02cc6c25eac0c77f858a93d2a; npm run release:check rerun on merged main passed (npm test 559/559, bundle/fixture/manifest/host-map/Codex adapter/review-drift/Claude plugin validation, npm publish dry-runs). @bbingz/polycli@0.6.28 was published by the maintainer during npm 2FA handoff; Codex published the missing @bbingz/polycli-opencode@0.6.28, created the GitHub release, and updated release docs.

2026-06-26 — Claude — provider-state review: live re-verify all 11 CLIs + drift fixes (PR #15, v0.6.28 release candidate)

  • Ran a provider-state-review Workflow (11 read-only probe agents → per-provider adversarial verify → synthesis, 23 agents) to re-check every provider CLI's live install + upstream + adapter contract on top of v0.6.27. Headline: no version gaps, no breaking CLI drift — all 11 locals == upstream where comparable (claude 2.1.193 / gemini 0.49.0 / qwen 0.19.2 / copilot 1.0.65 / opencode 1.17.11 / pi 0.80.2 / cmd 0.40.8 / mmx 1.0.16 / kimi-code 0.19.1 / grok 0.2.64 / agy 1.0.12); every flag/auth/argv assumption verified intact against live --help.
  • (code) copilot resume contract — buildCopilotInvocation now emits --session-id <id> instead of --resume <id> for resume-by-exact-id. copilot 1.0.65's -r, --resume[=value] takes an OPTIONAL =-attached value (or opens the session picker), so the prior space-separated form would not resume by id; --session-id <id> is the documented by-id flag. Reachable via ask/rescue --provider copilot --resume <id> (the companion sets resumeSessionId). Regression updated in copilot.test.js; all 5 companion bundles regenerated.
  • (code) minimax finish-reason — extractMiniMaxResponseFromMmxJson now honours Anthropic-style stop_reason in the finishReason fallback (mmx speaks both the OpenAI finish_reason and Anthropic Messages stop_reason shapes, and the parser already handles Anthropic content[] blocks). Pure additive, zero regression; minimax.test.js now asserts finishReason from the existing stop_reason fixture.
  • (docs/comments) kimi version-label refresh — kimi-code v0.6.0 → kimi-code 0.19.1 in docs/provider-paths.md + docs/polycli-v1-public-surface.md; dropped the re-drifting v0.6.0 pin from behavioural code comments (kimi.js, review.mjs, prompt-runtime.mjs) since the behaviour is version-general. Bumped the provider-paths.md snapshot to 2026-06-26 and replaced the stale v0.6.21 clause in the roadmap.md Current-state section with the 2026-06-26 re-verification note. (Project memory reference_cli_provider_versions.md rewritten to the 11-provider reality: pi @mariozechner→@earendil-works; kimi spawns kimi-code not the shadowed PyPI kimi-cli; minimax=mmx-cli not mini-agent; self-updating kimi/grok/agy have no read-only latest channel.)
  • Deferred / FLAGGED (not changed) per minimum-diff + the AGENTS.md "flag pre-existing dead code, don't delete" rule: pi.js dead agent_end.result.text/resultEvent.error branches (live agent_end carries only messages[]/willRetry, verified vs pi-agent-core .d.ts; harmless); the opencode.js billed run "ping" auth probe could use non-billing opencode auth list but auth-list proves CONFIGURED-not-WORKING, so folding it in would weaken status honesty / risk the four-state; and the JSON/stream-json event schema for ~7 providers stays unverified (read-only/cost constraint — needs an execution-allowed run or fixture recapture; gemini/grok fixture meta versions also lag, already tracked by check:fixture-freshness).
  • Validation: npm test 559/559, npm run release:check exit 0. Respects the Path B boundary (no shared runtime base, no parser promoted into polycli-utils, timing four-state untouched, cold/retry still unimplemented).

2026-06-19 — Claude — release: v0.6.27 published (review residual cleanup)

  • Published v0.6.27 to npm: @bbingz/polycli@0.6.27 (shasum 397b2349bd3c952c2b612c7f762a9db48e09cb09) and @bbingz/polycli-opencode@0.6.27 (shasum e38d544a851ad67302aa50b7f75d028b80cb6100), both latest. GitHub release v0.6.27 (publishedAt 2026-06-19T09:29:35Z) + tag v0.6.27. Utility packages unchanged.
  • Cleared the remaining v0.6.26-review residuals via PR #13 (Node 20 CI green, rebase-merged): fixed a background-job disk leak (saveState now reclaims the result/config/log artifacts of terminal jobs pruned past MAX_JOBS, wiring the previously-dead removeJobFile + a new removeJobLogFile); enforced the docs/capture-fixtures.md path/meta contract in validate-fixture-metadata.mjs (provider==dir, name==stem); added OpenCode exit-2 execution-path coverage (runCompanion exported with an injectable spawn); and synced the docs/roadmap.md Current-state section that still said the latest release was v0.6.24. A repo-wide scan confirmed no other release-state claim was stale. See docs/release-notes-v0.6.27.md.

2026-06-19 — Claude — release: v0.6.26 published (review fixes)

  • Published v0.6.26 to npm: @bbingz/polycli@0.6.26 (shasum b1ec2bcf366f1974e6850c42ca8c3ee81695999a) and @bbingz/polycli-opencode@0.6.26 (shasum f1e86227af994c281d0fe860c59809b04c103470), both latest. GitHub release v0.6.26 (publishedAt 2026-06-19T09:03:11Z) + tag v0.6.26. Utility packages unchanged.
  • Addresses an external review of v0.6.25 via PR #12 (Node 20 CI green, rebase-merged): (High) fixed extractTerminalError missing a nested error object ({error:{message:...}}) so visible partial text was wrongly ok:true — covers parseGrokJsonResult + parseGrokStreamText, with json/streaming/empty-object regressions; (Medium) constrained the cc-X validator status to verified/marketplace-unstable and clarified roadmap Q10 that the validator guards structure + source-anchoring, not current-truth; (Medium) synced release-state docs (README en/zh/ja + roadmap snapshot) that still said v0.6.24; (Low) added a runQwenPrompt --model argv regression. See docs/release-notes-v0.6.26.md.
  • Deferred (declared): validate-fixture-metadata.mjs path/meta consistency is a pre-existing Low gap, flagged for a separate change.

2026-06-19 — Claude — release: v0.6.25 published

  • Published v0.6.25 to npm: @bbingz/polycli@0.6.25 (shasum f243987016b6b89d536aadb83314a9416acd52e8) and @bbingz/polycli-opencode@0.6.25 (shasum 387fbf0347c5abc498d632c654b29783613240d5), both latest. GitHub release v0.6.25 (publishedAt 2026-06-19T08:09:43Z) + tag v0.6.25. Utility packages unchanged (@bbingz/polycli-utils@1.0.2, @bbingz/polycli-timing@1.0.1).
  • Bundles the three entries below (re-verification remediation + tmux test stabilization + cc-X endpoint recipes). Landed via PR #11 (Node 20 CI green; the previously-flaky tmux test passed on CI), merged to main via rebase. See docs/release-notes-v0.6.25.md.

2026-06-19 — Claude — docs: cc-X domestic-model endpoint recipes (Path-B docs + reference data)

  • Added docs/cc-x-endpoints.md (human reference) + docs/cc-x-recipes.json (machine-readable source of truth) encoding the cc-X pattern: point the EXISTING claude runtime (BYOK) or opencode (OpenAI-compatible) at a domestic vendor's Anthropic-compatible endpoint via ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN + ANTHROPIC_MODEL. Covers 9 entries across 7 PRC core labs (MiniMax, Moonshot Kimi, Zhipu GLM, Alibaba Qwen, DeepSeek, ByteDance Doubao, StepFun, Baidu Qianfan, Tencent) with per-vendor base URL, model-id family, native-CLI grouping, context-window (autoCompactWindow), caching note, and a source URL+date per entry.
  • Encoded the operational gotchas: silent prompt-cache degradation on shim endpoints (dual cache-breakpoint; DeepSeek is the auto-prefix-caching exception), pin a known-good Claude Code version + CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1, size CLAUDE_CODE_AUTO_COMPACT_WINDOW to the model's context, marketplace (Baidu/Tencent) model-identity instability, and the PRC data-sovereignty/Entity-List gate as SEPARATE from harness choice.
  • Honest-default: marketplace/resale endpoints carry status: "marketplace-unstable" and autoCompactWindow: null (no fabricated model/version pin), mirroring the gemini attempted-vs-used-model caveat in docs/model-fallback-policy.md. Enforced by scripts/validate-cc-x-recipes.mjs (a pure validator modeled on validate-fixture-metadata.mjs) + scripts/tests/validate-cc-x-recipes.test.mjs (auto-joined by the npm-test glob); added npm run validate:cc-x-recipes for standalone use.
  • Documented that cc-X is NOT a polycli provider/adapter/runtime — it rides the existing runtimes via standard env vars; the claude -p path forwards them via full process.env inheritance, while the tmux allowlist (CLAUDE_TMUX_ENV_EXACT) forwards the ANTHROPIC_* trio but NOT the CLAUDE_CODE_* knobs (documented, not fixed). Clarified that the polycli minimax/mmx-cli provider is a stateless text/media call, not the MiniMax cc-X coding path.
  • Cross-linked from docs/provider-paths.md (new subsection + Official-references bullet) and docs/polycli-v1-public-surface.md (one out-of-contract sentence). Recorded the no-adapter decision in docs/roadmap.md as closed Q10 + an Explicit-non-goals bullet. Zero runtime/production-path code change; claude.js env behavior left untouched by design. Verification: node scripts/validate-cc-x-recipes.mjs ok (9 entries), node --test scripts/tests/validate-cc-x-recipes.test.mjs 5/5, npm test + npm run release:check green. Snapshot facts are 2026-06-19; the validator guards structure + source-anchoring, not current-truth.

2026-06-19 — Claude — adversarial re-verification of the workflow-review remediation

  • Independently re-verified the committed remediation sweep (d272042 + 03ae92d) with a Workflow fan-out (9 adversarial auditors -> double-refutation -> completeness critic). 18 raw findings -> 7 confirmed + 1 critic-confirmed; 11 refuted. Confirmed the prior fixes are sound and re-ran full validation (the prior round's open residual #1): npm test 544/544, npm run release:check exit 0.
  • Closed residual #3 (state-root permissions) with real-filesystem evidence: under permissive umask 000, stateRoot/stateDir/jobsDir resolve to 0700 and state.json/job-config to 0600, enforced by explicit chmod (not umask). Characterized residual #4 (orphan <jobId>.json result files leak after MAX_JOBS pruning) as a PRE-EXISTING latent issue — removeJobFile is a dead export and the old code pruned identically — so it is out of scope for this remediation and left flagged, not fixed.
  • Fixed 2 confirmed regressions introduced by the remediation: (1) the opencode host adapter threw on exit code 2, but 2 is the companion's documented soft signal (health with no healthy provider, status --wait timeout) that still emits a valid JSON envelope on stdout — extracted isHardCompanionFailure(status) so exit 2 returns the envelope while exit 1/4/5/crash still reject; (2) cancelJob ran cleanupRuntimePaths (which deletes a review job's live cwd via cleanupPaths) BEFORE killing the worker — reordered to kill first, then clean up, and skip the runtime-path deletion entirely when the kill fails (worker may still be alive).
  • Fixed 2 confirmed incomplete fixes: (1) Grok SUCCESS_STOP_REASONS omitted MaxTokens, so a truncated-but-visible answer was wrongly marked ok=false — added maxtokens/max_tokens/length (grok-build's real StopReason enum is {EndTurn, MaxTokens, MaxTurnRequests, Refusal, ToolUse, Cancelled}, verified against the installed binary); refusal/cancelled/tool_use/max_turn_requests stay non-success; (2) the run-ledger append path created ~/.polycli/state/<slug> world-traversable (0o755) via the mode-less ensureParentDir on the run_started event that fires before any other state write — appendRunLedgerEvent now calls ensureStateDir first to land it 0o700.
  • Closed 4 confirmed test gaps (all mutation/RED-proven): pre-existing-0755 dir hardening test for ensureStateDir (state-1); state-dir-0700-after-append-only test for the run-ledger path (pwp-2, RED-proven); Grok non-success-stopReason-ALONE failure tests for both parseGrokJsonResult and runGrokPromptStreaming plus a MaxTokens-success test (test-1 + grok-1, RED-proven); sync runProviderPrompt explicit-model-before-default fallback test mirroring the streaming case (qwen-model-1); new scripts/tests/opencode-host.test.mjs pinning the exit-2 soft-signal contract (oc-status-1).
  • All changes respect the Path B architecture boundary: no shared runtime base class, no provider parser promotion into polycli-utils, timing four-state untouched, cleanupPaths still sourced only from internal review temp dirs.
  • Verification: focused RED/GREEN proofs for grok-1 and pwp-1 (reverting each fix turns its new test red); focused suite 66/66; npm test 544/544 (535 + 9 new tests); npm run release:check exit 0 (plugin bundles 5, fixture metadata 17, codex adapter 5; one tmux.jsonl ENOENT flake on the first run was the known full-suite-parallel-load flake — claude.test.js passes 28/28 in isolation, and the re-run was clean). Not published; current unreleased workspace work after v0.6.24.

2026-06-16 — Codex — Grok fixture residual cleanup

  • Closed the remaining workflow-review residual risk by capturing a real Grok streaming fixture with grok 0.2.51 (f4f85a6492e) [stable]: grok -p 'Reply with exactly HELLO_GROK_FIXTURE and nothing else.' --output-format streaming-json -m grok-build --permission-mode plan --disable-web-search --max-turns 1.
  • Added packages/polycli-runtime/test/fixtures/grok/stream-success.* and wired Grok into the table-driven fixture replay test, so Grok's real thought / text / end streaming shape is now parser-checked in CI.
  • Removed the default Grok missing-success allowlist from validate:fixtures; the command now checks 17 fixture metadata files and prints no missing-success allowlist rows.
  • Updated the hardcoded Grok fallback/default-model guidance from stale grok-composer-2.5-fast to current local default grok-build, and aligned the Grok plugin guidance skills plus fixture-capture docs.
  • Verification: node --test packages/polycli-runtime/test/grok.test.js packages/polycli-runtime/test/fixture-replay-all.test.js scripts/tests/validate-fixture-metadata.test.mjs && node scripts/validate-fixture-metadata.mjs passed; node --test packages/polycli-runtime/test/*.test.js scripts/tests/validate-fixture-metadata.test.mjs && node scripts/validate-fixture-metadata.mjs && node scripts/check-fixture-freshness.mjs passed with Grok fresh and older provider fixture-staleness warnings remaining warn-only; npm test passed 535/535; npm run release:check exit 0, including validate:fixtures at 17 checked and no allowlist output. Not published.

2026-06-16 — Codex — workflow review remediation sweep

  • Remediated the confirmed medium/low findings from the workflow deep review batch without changing the Path B architecture: no shared provider base class, no provider parser promotion into polycli-utils, and no timing-state collapse.
  • Hardened local privacy defaults: terminal fallback state now uses ~/.polycli/state instead of shared OS temp; state/workspace/job dirs are created private; state, job config/result, ledger, timing, preview, and provider model cache writes use private modes where applicable.
  • Fixed background job lifecycle gaps: cancellation now records cancelled terminal ledger events, removes per-job config, cleans runtime cleanupPaths, and MAX_JOBS pruning preserves queued/running jobs. The queued-to-running parent write now uses a stale-write guard so an already-terminal worker result is not overwritten.
  • Fixed host/provider semantics: OpenCode host adapter treats every non-zero companion exit as failure even with stdout; Qwen forwards explicit --model; registry model fallback now prefers provider output, then explicit model, then cached/default model; Grok marks terminal error metadata or non-success stop reasons as failed while preserving partial text.
  • Closed observability/test/doc gaps: TUI recognizes provider_decision:passed, all captured runtime fixtures are replayed through provider parsers, fixture metadata requires success fixtures for parser-backed providers, sessions list/purge has companion wiring integration coverage, README/public-surface/host-map/docs drift was aligned, and CI now dry-runs the terminal package tarball.
  • Stabilized the Claude tmux TUI fake-bin tests under full-suite parallel load by widening the test-only tmux timeout budget; this fixes the tmux.jsonl ENOENT flake seen during full-suite verification.
  • Verification: focused RED/GREEN tests for permissions, cancel cleanup, active job pruning, OpenCode non-zero exits, Qwen model forwarding, Grok terminal errors, TUI passed status, fixture replay/metadata, sessions CLI wiring, and queued-to-running stale writes; node --test packages/polycli-runtime/test/claude.test.js passed 28/28 after test stabilization; npm test passed 535/535; npm run release:check exit 0, including bundle/fixture/manifest/host-map/Codex adapter/review-drift checks, Claude plugin validation, and npm pack dry-runs for opencode/utils/timing/terminal.
  • Follow-up note: the remaining Grok real-fixture gap was closed in the later Grok fixture residual cleanup entry above. Not published; this is current unreleased workspace work after v0.6.24.

2026-06-16 — Codex — post-v0.6.24 latest-package review cleanup

  • Ran @bbingz/polycli@0.6.24 from npm against the v0.6.23..HEAD release diff: health --json found gemini, qwen, minimax, claude, copilot, opencode, pi, cmd, agy, and grok healthy; kimi remained quota-blocked with 403.
  • Dispatched independent background reviews through claude, copilot, opencode, pi, cmd, gemini, qwen, minimax, and grok. All 9 completed; agy review correctly rejected because agy cannot enforce non-interactive read-only plan mode. Five providers reported no issues; four raised low/medium candidates.
  • Adjudication: overturned package-lock drift because root workspaces are only packages/*, so plugins/polycli-opencode has no lockfile package entry; overturned stale docs/release-notes-v0.6.23.md latest wording as historical release-note state, not current release state.
  • Fixed the confirmed compatibility regression where status --all --timeout-ms abc --json failed without --wait; timeout values are now parsed only for wait paths, preserving the previous no-wait status behavior.
  • Hardened regression coverage: single-job status <jobId> --wait --timeout-ms 1 --json now asserts exit code 2 and waitTimedOut:true; invalid timeout coverage now checks both all-job and single-job wait paths; the fake delayed job timeout test uses a wider 3000ms delay to avoid slow-runner flakes.
  • Verification: focused RED/GREEN test for status wait timeout cases; node --test plugins/polycli/scripts/tests/integration.test.mjs passed 60/60; npm test passed 516/516; npm run release:check exit 0, including bundle/fixture/manifest/host-map/Codex adapter/review-drift checks, Claude plugin validation, and npm pack dry-runs. Not published; this is current unreleased workspace work after v0.6.24.

2026-06-16 — Codex — status wait timeout hardening

  • Ran a second real multi-provider review of the v0.6.23 release surface. Reviews completed through claude, copilot, opencode, pi, cmd, gemini, qwen, minimax, and grok; kimi remained quota-blocked; agy correctly rejected /review because it cannot enforce a read-only non-interactive plan mode.
  • Fixed confirmed status --all --wait timeout findings: JSON mode now exits 2 on timeout, text mode prints Timed out waiting for all jobs., invalid --timeout-ms values are rejected as positive-integer errors, and the all-job waiter no longer performs an unused initial snapshot read.
  • Applied the same timeout parser and timeout exit-code handling to the existing single-job status --wait path.
  • Hardened the explicit Claude tmux TUI worker regression assertion against missing timing.meta.
  • Verification: focused red/green regressions; node --test plugins/polycli/scripts/tests/integration.test.mjs passed 60/60; npm test passed 516/516; npm run release:check exit 0, including bundle/fixture/manifest/host-map/Codex adapter/review-drift checks, Claude plugin validation, and npm pack dry-runs.
  • Published GitHub release v0.6.24: https://github.com/bbingz/polycli/releases/tag/v0.6.24 (publishedAt 2026-06-16T07:26:49Z).
  • npm latest verified as 0.6.24 for both public host packages: @bbingz/polycli-opencode@0.6.24 (time.modified 2026-06-16T07:28:01.606Z, shasum 5da8640b1bba6b3da6309bd87692596c9cc8fb34) and @bbingz/polycli@0.6.24 (time.modified 2026-06-16T07:28:13.403Z, shasum 8a766b320a3f5ed18b6e083ab98b87c6fc753b9e).

2026-06-16 — Codex — post-release full-provider smoke fixes

  • Ran a real Polycli full-provider smoke review of the v0.6.22 diff (--base v0.6.21 --scope branch) and adjudicated provider reports against source. The release diff itself had no confirmed correctness/security regression; the smoke exposed two Polycli control-plane bugs.
  • Fixed health --provider opencode: health probes now hydrate provider runtime env before calling runProviderPromptStreaming, preserving PATH when prompt constraints inject OPENCODE_CONFIG_CONTENT. This removes the false spawn opencode ENOENT result while keeping the deny-all opencode config in place.
  • Fixed status --all --wait: the command now waits for every active job and returns an all-job status snapshot instead of waiting only for the latest active job and returning a single-job envelope.
  • Added companion-level coverage for the explicitly retained Claude executionMode: "tmux-tui" worker path, plus regressions for opencode health env hydration and status --all --wait.
  • Verification: focused red/green regressions; live polycli health --provider opencode --json returned healthy; node --test plugins/polycli/scripts/tests/integration.test.mjs passed 58/58; npm test passed 514/514; npm run release:check exit 0, including bundle/fixture/manifest/host-map/Codex adapter/review-drift checks, Claude plugin validation, and npm pack dry-runs.
  • Published GitHub release v0.6.23: https://github.com/bbingz/polycli/releases/tag/v0.6.23 (publishedAt 2026-06-16T06:44:46Z).
  • npm latest verified as 0.6.23 for both public host packages: @bbingz/polycli-opencode@0.6.23 (time.modified 2026-06-16T06:49:58.445Z, shasum 96a99bb18f69fd40dd8a3c78506311fc89b0d0d7) and @bbingz/polycli@0.6.23 (time.modified 2026-06-16T06:50:22.282Z, shasum 02d016850b5998eabb2bb3faefa6c12ca7e4bfcc).

2026-06-16 — Codex — v0.6.22 published (Claude print defaults)

  • Anthropic paused the Agent SDK / claude -p dedicated-credit change, so the previous default tradeoff no longer holds for ordinary Claude ask / review.
  • Restored Claude ask / review defaults to headless claude -p while preserving plan/no-tools/no-MCP constraints (--permission-mode plan --tools "" --mcp-config '{"mcpServers":{}}' --strict-mcp-config).
  • Kept the detached tmux TUI runtime path intact for explicit/internal callers, especially workflow cases that need an interactive Claude Code runtime.
  • Updated focused tests and current docs to separate the historical v0.6.21 tmux-default release from the current main behavior.
  • Live smoke: POLYCLI_TMUX_BIN=/tmp/polycli-no-tmux-for-print-smoke node plugins/polycli/scripts/polycli-companion.mjs ask --provider claude --json ... returned ok:true, response POLYCLI_CLAUDE_PRINT_SMOKE_20260616, measured ttft/gen/tail, session e639c2cb-320f-4226-b521-ed5b608851b9, and Claude-reported total_cost_usd:0.11984500000000001.
  • Published GitHub release v0.6.22: https://github.com/bbingz/polycli/releases/tag/v0.6.22 (publishedAt 2026-06-16T02:52:57Z).
  • npm latest verified as 0.6.22 for both public host packages: @bbingz/polycli-opencode@0.6.22 (time.modified 2026-06-16T02:51:05.698Z, shasum 09e36dbd10d2bc72257f3c27ed3b6b910809901e) and @bbingz/polycli@0.6.22 (time.modified 2026-06-16T02:51:14.262Z, shasum 28b00344f743ec0b37342242c80b81867b293c73).

2026-06-15 — Codex — Claude workflow orchestration design

  • Researched the current Claude Code Dynamic Workflow surface, Codex xhigh planning options, and local Claude/Qwen/Minimax/Kimi/MiMo workflow run artifacts to choose an implementation path for multi-level subagent workflows.
  • Added docs/superpowers/specs/2026-06-15-claude-workflow-orchestration-design.md. The design keeps Codex xhigh as a planner/compiler for workflow JS, uses Claude Code Dynamic Workflows as the actual subagent runtime, and launches Claude through the existing tmux TUI path so the default does not fall back to claude -p or Agent SDK credit usage.
  • Scoped polycli to a thin control/observability surface (workflow plan/start/list/status/result) instead of a new agent framework. Workflow artifact readers must treat Claude's workflows/wf_*.json and subagents/workflows/<id>/*.jsonl format as observed local evidence, not a guaranteed public API.

2026-06-15 — Codex — v0.6.21 published (Claude tmux TUI + review remediation)

  • Released v0.6.21 after the third-party review remediation closeout. GitHub release/tag to be created from this release commit; npm publication is already visible on the registry.
  • npm latest verified via npm view: @bbingz/polycli@0.6.21 and @bbingz/polycli-opencode@0.6.21. Utility packages remain on their independent v1 line: @bbingz/polycli-utils@1.0.2, @bbingz/polycli-timing@1.0.1.
  • Release content: Claude ask/review default to detached tmux TUI mode instead of claude -p; tmux launch responses expose tmuxSession/attachCommand and startup-only timing; review gate sentinel parsing, auth probing, fixture freshness, state cleanup, tempfile cleanup, unsafe pid guards, and docs parity are hardened.
  • Verification: npm run release:check exit 0 (npm test 511/511 + bundle/fixture/manifest/host-map/Codex adapter/review-drift/Claude plugin validation + npm dry-runs). npm whoami authenticated as bbingz; npm publish completed for the two v0.6.21 public host packages.

2026-06-15 — Codex — final multi-review cleanup before merge closeout

  • Re-ran a source-grounded adjudication over the remaining non-Qwen review findings after PR #9 landed. One additional behavior bug was confirmed still-present: the stop-time review gate scanned all response lines for bare ALLOW: / BLOCK: sentinels, so a provider echo of the previous Claude response could be misread as the gate verdict.
  • Fixed the stop-review gate by generating a per-run POLYCLI_STOP_REVIEW_* token, requiring ALLOW <token>: / BLOCK <token>: in the provider response, and ignoring stale bare sentinels when the token is active. Added parser and runStopReview regressions.
  • Rechecked the old isTerminalSummaryEvent finding. It is not a live MiniMax/cmd/agy/kimi bug in the current runtime: MiniMax declares ttft/tail unsupported, cmd/agy stream only text-delta events, and Kimi's meta event carries no visible text. No code change kept for that claim.
  • Remaining non-bug findings are tracked as design/maintenance items, not release blockers: Claude ask/review stays detached tmux TUI by product requirement; Claude health stays auth-only; provider env filtering and duplicated transient-pattern helpers are future hardening/refactor candidates.

2026-06-15 — Codex — Qwen audit checklist remediation

  • Consolidated the Qwen third-party review batch into docs/audit/third-party-review-followup-2026-06-15.md, then verified all 11 claims with independent subagents against the current worktree before editing. All 11 were confirmed still-present before remediation.
  • Fixed the three behavior/security issues with regressions: writeFileAtomicSync now removes its temp file on failed rename/write paths; no-diff review cleanup now runs in a finally so Gemini isolated tempdirs are removed; unsafe pid values (<=1 / non-integers) are rejected before process-group termination.
  • Added the missing Claude health logged-out integration coverage by making the fake Claude auth fixture emit loggedIn:false via env, and asserting the companion reports Claude unhealthy with a populated probe error.
  • Closed the docs/parity findings: plugin README lists debug / sessions and terminal TUI ownership; root and translated READMEs describe Claude health as auth-only, add the terminal package badge, outcome diagnostics, and minimax (mmx-cli) alias; timing/runtime package READMEs document v1 cold/retry and REVIEW_FLAG_EXPECTATIONS.
  • Verification: focused regressions pass; npm test exit 0 (508/508); npm run release:check exit 0 including bundle, fixture, manifest, host-map, Codex adapter, review-drift, Claude plugin validation, and npm pack dry-runs.

2026-06-15 — Codex — multi-review adjudication follow-up

  • Adjudicated the Minimax/Kimi/MiMo review batch against the current source after the Claude tmux TUI remediation. Kept the user-requested Claude tmux TUI default instead of reverting ask/review to claude -p; treated "restore synchronous LLM answer" findings as a product-semantics conflict, not a fix to apply.
  • Fixed two confirmed issues: Claude legacy auth status non-JSON success output is now parsed or marked inconclusive instead of treated as logout, and session-lifecycle-hook.mjs now removes session jobs through locked updateState rather than naked load/save.
  • Closed release-safety/doc drift found in the review batch: fixture freshness probes now cover the 11-provider runtime surface (cmd, agy, grok included); README capability notes, docs/provider-paths.md, docs/polycli-v1-public-surface.md, CLAUDE.md, and docs/roadmap.md describe Claude tmux TUI startup-only timing and the tmuxSession/attachCommand response shape.
  • Added draft docs/release-notes-v0.6.21.md for the current unreleased patch rather than rewriting the already-published v0.6.20 notes.

2026-06-14 — Codex — Claude tmux TUI review remediation

  • Adjudicated the Claude/DeepSeek review findings against the current code and the user requirement that Claude subagent calls avoid the upcoming claude -p pay-as-you-go path. Confirmed the ask/review semantic drift, timing ambiguity, missing signal cleanup, tmux environment propagation gap, and auth-only health ambiguity; intentionally did not revert Claude ask/review defaults to -p.
  • Hardened Claude tmux TUI mode: tmux new-session now receives an explicit allowlist of Claude/Anthropic/proxy/cert env vars via -e; SIGINT/SIGTERM during orchestration kill the created tmux session before process shutdown; missing tmux gets a direct install/config error; successful tmux launches return detached:true, responseKind:"tmux_tui_session_started", warnings, and timingMeta that says timing covers only tmux_startup and llmCompletionObserved:false.
  • Runtime timing now merges provider timingMeta and uses the run-level timing support for Claude tmux TUI, so ttft/gen/tail stay unsupported, total remains schema-valid measured, and the record explicitly marks tmuxDetached:true / startup-only timing. Claude health remains no-model-call/auth-only by design and now reports probe.kind:"auth_status" plus authOnly:true instead of looking like a sentinel LLM probe.
  • Tests added/updated for tmux env propagation, detached payload semantics, startup-only timing metadata, signal cleanup, Claude health auth-only reporting, and companion ask/review integration. Bundles regenerated for all host surfaces.
  • Verification: npm test exit 0 (500/500); node --test packages/polycli-runtime/test/claude.test.js; node --test packages/polycli-runtime/test/registry.test.js; node --test plugins/polycli/scripts/tests/integration.test.mjs; npm run validate:bundles; npm run validate:manifests; npm run validate:host-map.

2026-06-02 — Claude — repo cleanup: removed stale R8 worktrees + release/v0.6.19 branch

  • After the v0.6.20 release, deleted the merged release/v0.6.19 branch and the 3 abandoned worktree-agent-* git worktrees + their branches (all local-only — none on origin). Verified safe first: each branch had 0 commits not in main (so git branch -d succeeded, git-confirming they were merged); the worktrees' only uncommitted content was an identical, obsolete 2026-04-24 path-rewrite (/home/user/…→<local-home>/…) on a snapshot ~41k lines behind main, locked by a dead pid (96484).
  • The single (identical across all 3) staged diff was saved to /tmp/r8-worktree-staged-pathrewrite.patch as insurance, but applying it is NOT advised: active files (README/docs) no longer carry those paths, and the remaining /home/user/ references on main are historical records (CHANGELOG, docs/archive/*, release-notes-v0.6.1) that should not be rewritten.
  • Repo now has a single main branch, synced with origin, at v0.6.20.

2026-06-02 — Claude — docs: kimi-code v0.6.0 skill/docs refresh + README 11 providers (PR #8)

  • Closed the doc-debt deferred through v0.6.20 (merged via PR #8). The kimi skill prose + reference docs still described the legacy Python kimi-cli, contradicting the kimi-code v0.6.0 adapter: rewrote kimi-cli-runtime/SKILL.md (-p one-shot, structured session.resume_hint id, --session/-C resume, ~/.kimi-code/, prompt-only review); fixed kimi-prompting/SKILL.md rule 4 (--max-steps-per-turn is config-level now), and the kimi rows in docs/provider-paths.md + docs/polycli-v1-public-surface.md.
  • README (GitHub-rendered) provider prose list, Hosts/providers table, and capability matrix now list all 11 providers — added the previously-missing agy (since v0.6.18) and grok (v0.6.20); readme-header.svg "ten"→"eleven".
  • Docs-only (no code/bundle change). npm test 483/483; validate:codex-adapter / host-map / manifests green.

2026-06-02 — Claude — v0.6.20 published (GitHub release + npm)

  • Released the merged grok / kimi-code / deep-review work as v0.6.20. GitHub release: https://github.com/bbingz/polycli/releases/tag/v0.6.20.
  • npm latest (verified via npm view): @bbingz/polycli@0.6.20, @bbingz/polycli-opencode@0.6.20, @bbingz/polycli-utils@1.0.2 (bumped for the atomic-save/process/stream fixes); @bbingz/polycli-timing@1.0.1 unchanged.
  • Gate: npm run release:check exit 0 (npm test 483/483 + validate:bundles/fixtures/manifests/host-map/codex-adapter + check:review-drift + claude plugin validate ×2 + publish dry-runs). See the gate/merge entry below for the substantive changes.

2026-06-02 — Codex — baseline GitHub branch protection for public non-fork repos

  • Configured minimal default-branch protection for all bbingz public, non-fork repositories that currently have a default branch: polycli, engram, Rules, kimi-plugin-cc, gemini-plugin-cc, minimax-plugin-cc, qwen-plugin-cc, tech-microgpt-cn, ZhaoShengList, feishu-chatgpt, bbingz.github.io, and gfw.
  • The protection baseline blocks force pushes and branch deletion, and applies to admins (protected=true, allow_force_pushes=false, allow_deletions=false, enforce_admins=true). It intentionally does not require status checks or PR reviews, so normal push/merge behavior is not gated by this change.
  • Verification: listed public repos with gh repo list bbingz --visibility public --limit 1000, wrote branch protection with gh api --method PUT repos/{repo}/branches/{default}/protection, and read back each default branch plus protection object. bbingz/bingz has no default branch, so it was not applicable; public forks were not modified.

2026-06-02 — Claude — Codex review-gate on PRs #5/#6/#7, then merge all three to main (11th provider grok; unreleased)

Three independent PRs (deep-review hardening, kimi→kimi-code v0.6.0 migration, grok provider) went through a pre-merge Codex review gate, were fixed where findings were real, and were merged to main in order #5→#6→#7. NOT released — latest published release is still v0.6.19; this work accumulates for a future release.

Gate adjudication (Codex is not ground truth — every finding was checked against the real code):

  • PR #7 (grok) — Codex CHANGES_REQUESTED, all 3 real: (a) auth-probe ordering bug — /\blogged in\b/i banner check ran before the explicit auth-error regex, so a logged-out not logged in (which contains the substring logged in) was misread as loggedIn:true; reordered + regression test. (b) grok was half-wired — added it everywhere agy is advertised (polycli-codex + polycli-copilot skill descriptions/arg-hints, polycli-codex README + plugin.json, root README, host-command-map, and validate-codex-adapter PROVIDERS). (c) grok-cli-runtime SKILL over-claimed --effort mapping (polycli --effort is gemini-only, dropped for grok).
  • PR #6 (kimi-code) — Codex CHANGES_REQUESTED; 1 real, reframed: buildKimiInvocation emitted -r <id> for resume-by-id, but kimi-code v0.6.0 has no -r (kimi --help: resume-by-id is -S, --session [id], continue-last -C); the path is reachable via rescue --resume <id>, so -r would be rejected at runtime — switched to --session. Also tightened the session-id parse to require type==='session.resume_hint'. Fixed a fragile review integration test exposed by committing the migration (doesNotMatch(argv.join()) also matched the reviewed diff text embedded in the -p prompt → false positive; now checks flags as discrete argv tokens). Codex false-positives (verified, not changed): the -p+--plan/--auto/--yolo combination is latent-only (no caller injects those), and the ~/.kimi/ literal is an intentional migration-history comment.
  • PR #5 (deep-review) — Codex review stalled (~20 min, no output) on the 1249-line diff, so the gate ran a Claude-driven 6-dimension adversarial workflow (20 reviewers): auth-probe transient cluster, atomic-save locking, signal-kill, stream limit, job-control concurrency, state dedup, companion sessionId, test adequacy. 14 raw findings → 0 survived adversarial verification (the 25 hardening fixes are correctly implemented, no regressions). Only actionable item: untracked an accidentally-committed .codegraph/.gitignore and added .codegraph/ to root .gitignore.

Merge mechanics: source 3-way merges were clean (provider entries from kimi + grok both preserved; PROVIDERS includes grok). The 5 polycli-companion.bundle.mjs files conflicted/auto-merged textually but git's textual bundle merge did NOT match the source — regenerated all bundles via npm run build:plugins so they are byte-identical to the merged source.

Verification on merged main: npm test exit 0 (483/483, up from 453); validate:host-map (12 capabilities), validate:codex-adapter (5 files, now includes grok), check:review-drift (no drift) all exit 0; PROVIDER_IDS = 11. docs/roadmap.md updated to 11-providers-in-main / unreleased. NOT run: release:check + npm publish (no release this round).

2026-05-30 — Codex — review v0.6.19 upgrade docs for current-state drift

  • Reviewed the v0.6.19 upgrade range (v0.6.18..HEAD) after Claude's maintenance/session-pollution increment. No code-level regressions were found in the new review-flag, session-artifact, bundle, or release-check paths under current verification.
  • Fixed two durable documentation facts that had drifted after the release closeout: docs/roadmap.md now reports v0.6.19 / 10 providers as current state and lists the new fixture-freshness guardrail; docs/archive/session-memory-2026-05-30.md now records 84621b1 as the publish closeout commit instead of a permanently current main HEAD.
  • Verification: npm test exit 0 (453/453); focused new-area tests exit 0 (69/69); npm run validate:bundles, validate:fixtures, validate:manifests, validate:host-map, and validate:codex-adapter exit 0; npm run check:review-drift exit 0 (all installed CLI flag probes ok); npm run check:fixture-freshness exit 0 with 16 expected STALE warnings; claude plugin validate passed for both Claude manifests; git diff --check v0.6.18..HEAD exit 0.

2026-05-29 — Claude — v0.6.19: Q8a/b/c maintenance hardening + Q9a/b upstream session-pollution control

Spec-driven increment (docs/superpowers/specs/2026-05-29-maintenance-and-pollution-design.md) from the 2026-05-29 strategy recon (memory project_competitive_landscape_and_moat, roadmap Q8/Q9). Two Codex review gates: spec → CHANGES_REQUESTED → rev2 (review 019e73b4); implementation → CHANGES_REQUESTED → fixes → APPROVE (aeec4314). Implemented by parallel sub-agent waves, integrated + verified by Claude. Published 2026-05-29: GitHub release v0.6.19 (https://github.com/bbingz/polycli/releases/tag/v0.6.19) + npm @bbingz/polycli-opencode@0.6.19 and @bbingz/polycli@0.6.19 (both latest, verified via npm view).

Q9a/Q9b — upstream session-pollution control (the user's #3 pain: spawned-CLI session files accumulate under ~/.claude, ~/.gemini, …):

  • run-ledger events now carry sessionId (between model and defaultModel) and a verified sessionArtifactPath (after sessionId), threaded at the companion foreground + worker run sites and the job-control recovery path where result/cwd/provider are in scope; null (never fabricated) where not (run-ledger.mjs, polycli-companion.mjs, job-control.mjs).
  • plugins/polycli/scripts/lib/sessions.mjs (NEW): deriveSessionArtifactCandidate (ONE exact path per provider, NO glob — claude ~/.claude/projects/<cwd '/'→'-'>/<id>.jsonl verified against the live store; kimi ~/.kimi/sessions/<md5(cwd)>/<id>/ dir; pi/gemini/codex/minimax/cmd → null+reason), recordArtifactPath (records only if exists + not a symlink + realpath under the provider store root), collectRecordedArtifacts, collectNonPurgeableSessions, planPurge (re-validates lstat/realpath/basename), executePurge (dry-run default; deletes only with --confirm).
  • new polycli sessions [list | purge --confirm] command + plugins/polycli/commands/sessions.md, registered across all host surfaces (validate-host-command-map + docs/host-command-map + Codex/Copilot skills + OpenCode index): 12 capabilities. Deletion is driven ONLY by recorded + re-validated realpaths — never path-guessing or globbing. Non-purgeable tracked sessions (gemini per-project dir, pi timestamp-prefixed filenames, ephemeral providers) are reported with a reason, never silently dropped.

Q8a/Q8b/Q8c — provider-drift maintenance hardening (root cause: ecosystem heterogeneity + duplication + version-pinned fixtures giving false confidence):

  • Q8a scripts/check-fixture-freshness.mjs (NEW; npm run check:fixture-freshness): warns when a fixture's pinned CLI version ≠ the locally-installed CLI (-v/-V/--version per provider), skips absent CLIs, exit 0 default / --strict non-zero on real staleness. On this machine all 16 fixtures are STALE (real upstream drift, e.g. claude 2.1.117→2.1.156, gemini 0.38.2→0.43.0) — exactly the silent false-confidence this catches.
  • Q8b single frozen REVIEW_FLAG_EXPECTATIONS map (packages/polycli-runtime/src/review-flags.js, NEW) is the sole declaration of each provider's drift expectFlags/forbidFlags/probes, read-only option key/value, and exact extraArgTokens. check-review-cli-drift.mjs derives its CHECKS from it; review.mjs sources read-only keys from it; a consistency test asserts extraArgTokens EXACTLY equals the --flags REVIEW_HARD_CONSTRAINTS emits (catches a token ADDED or REMOVED — the original subset check missed gemini/kimi entirely). Data co-location, NOT a BaseProvider (non-goal #1 intact).
  • Q8c check:review-drift wired into release:check (self-skips absent CLIs; blocks a release only on genuine flag drift) + a LOCAL regex-anchor sanity check reading the GEMINI_EXPLICIT_AUTH_ERROR_RE/KIMI_EXPLICIT_AUTH_ERROR_RE source (guards a polycli-side refactor from dropping the invalid api key anchor; does NOT detect upstream wording — a real upstream auth-wording probe stays an open follow-up).

Verification: npm test 453/453 (from 399); npm run release:check exit 0 (5 bundles byte-identical; fixture metadata 16; host-map 12 capabilities; codex-adapter; no CLI drift; npm pack/publish dry-runs). Deferred to roadmap, NOT shipped: Q8d (migrate churn-heavy providers to JSON/SDK — multi-release) and Q9c (opt-in env session isolation — a naive HOME/XDG override breaks auth + --resume; needs per-provider design).

2026-05-27 — Claude — v0.6.18 agy provider review fixes

  • Ran 5 review rounds on the agy provider (a836fa1..HEAD). Fixed the confirmed and clearly-actionable findings:
    • No fabricated session id (headline): agy.js no longer feeds result.stdout to resolveSessionId. agy stdout is pure assistant prose, so the UUID scan would capture any UUID in an answer (e.g. "give me a uuid") as a fake sessionId, violating the spec ("sessionId always null / do not fabricate") and suppressing buildTimingMeta's sessionIdMissing:true. sessionId is now hard-null on both sync and streaming paths; dropped the unused resolveSessionId import.
    • Hardened auth probe: buildAgyAuthStatus now inspects combined error+response text (catches a logged-out agy that prints sign-in guidance to stdout and exits 0 → loggedIn:false) and treats a clean status:0 with no auth signal as authenticated even when the probe produced no visible text (fixes the empty-output false-negative where the hasVisibleText gate leaked into auth classification). Transient→inconclusive-authenticated behavior preserved.
    • Review hints corrected: removed agy from the --provider argument-hint in commands/review.md and commands/adversarial-review.md; the runtime rejects agy review (host-command-map already marked it unsupported), so the hints no longer advertise an unsupported choice.
    • Drift watcher actually watches: check-review-cli-drift.mjs agy row had expect:[], a no-op that could only detect expected flags disappearing — never a new plan flag appearing, which is the row's whole purpose. Added a forbid list (--approval-mode/--permission-mode/--policy/--plan/--agent); the checker now reports DRIFT if any appear so /review support can be re-evaluated.
    • Tests: added 5 agy regression cases (UUID-in-output→null sessionId; empty-output→no_visible_text; authed-empty→loggedIn; logged-out-to-stdout-exit0→loggedOut; streaming non-zero→auth failure).
  • Verification: node --test packages/polycli-runtime/test/agy.test.js exit 0 (18/18); npm test exit 0 (399/399, up from 392); npm run release:check exit 0; node scripts/check-review-cli-drift.mjs runs clean ([ ok ] agy); all 5 companion bundles rebuilt and byte-identical (validatePluginBundles green). Published: tag v0.6.18, GitHub release, npm @bbingz/polycli-opencode@0.6.18 and @bbingz/polycli@0.6.18 (both latest, verified via npm view).
  • Deferred (not fixed, with rationale): signal-kill status:null→0 misclassification is a repo-wide runCommand pattern (all providers) — fix cross-cutting, not agy-only; stdout banner/notice pollution of response+ttft and the internal---print-timeout-beats-outer-timeout classification both depend on agy's actual print-mode output shape, which upstream has not been verified; response-vs-events blank-line divergence only affects preview events (compacted away). The v0.6.16 release notes' --add-dir/--sandbox mention describes runtime params with no companion CLI surface (historical, left as-is). validate-codex-adapter prompts≤3/≤128 is an intentional Codex limit, not a bug.

2026-05-25 — Codex — v0.6.17 Codex manifest prompt-limit patch

  • Fixed the Codex host manifest noise found in codex-tui.log: plugins/polycli-codex/.codex-plugin/plugin.json had 4 interface.defaultPrompt entries, while Codex currently supports a maximum of 3. The manifest now keeps health, ask, review, and timing coverage in 3 supported prompt entries.
  • Hardened scripts/validate-codex-adapter.mjs so release validation rejects more than 3 Codex default prompts and rejects any default prompt entry over 128 characters. Added focused regression tests for both limits.
  • Verification before release prep: node --test scripts/tests/validate-codex-adapter.test.mjs exit 0 (4/4); node scripts/validate-codex-adapter.mjs exit 0; node --test scripts/tests/*.test.mjs exit 0 (45/45). Release verification: npm run release:check exit 0 (394/394 tests; bundles 5; fixtures 16; manifests 0.6.17; host-map 11x4+terminal; codex-adapter 5; claude plugin validate ×2; npm dry-runs/pack checks passed). Post-publish: tag v0.6.17, GitHub release (not draft, not prerelease), and both npm packages observable at 0.6.17 (@bbingz/polycli-opencode@0.6.17 at 2026-05-25T12:01:21Z, @bbingz/polycli@0.6.17 at 2026-05-25T12:01:32Z).

2026-05-20 — Claude+Codex — add agy (Google Antigravity CLI 1.0.0) provider

  • Added agy as the tenth polycli-managed provider CLI. Adapter mirrors the text-only cmd pattern with claude-style session flags (--continue/--conversation <id>/--add-dir/--sandbox) and YOLO via --dangerously-skip-permissions. agy emits plain stdout (no JSON envelope, no session id, no model field); the adapter honors the four-state timing semantics by returning null model and the resolver's null sessionId rather than fabricating values.
  • TIMING_SUPPORT: ttft/gen/tail true (line-buffered stdout), tool false, runtimePersistence: "session". The session id is always missing, so buildTimingMeta will correctly stamp sessionIdMissing: true on every agy run — honest, not folded into unsupported.
  • /review --provider agy is rejected upfront via a new REVIEW_UNSUPPORTED_PROVIDERS set and assertReviewProviderSupported in plugins/polycli/scripts/lib/review.mjs. Rationale: agy has no plan-mode / approval-mode flag, so the review hard constraint cannot enforce read-only execution. Drift watcher (scripts/check-review-cli-drift.mjs) carries the agy row with expect: [] so future plan-mode additions are picked up.
  • ask/rescue auto-YOLO for agy (matches the project_yolo_standard.md rule), provider listings in companion dispatcher / Claude commands / Codex+Copilot skills / README SVG header updated from "nine providers" to "ten providers".
  • Verification: node --test packages/polycli-runtime/test/agy.test.js exit 0 (13/13); npm test exit 0 (392/392, up from 374); npm run check:provider-paths exit 0 (8 ok + agy ok + pi skipped on local timeout); npm run release:check exit 0 (bundles 5; fixtures 16; manifests 0.6.16; host-map 11x4+terminal; codex-adapter 5; claude plugin validate ×2; npm pack/publish dry-runs passed). agy itself reviewed commit a836fa1 and returned VERDICT: PASS against four-state-timing / YOLO / Path B / review-refusal / plain-text-stdout invariants. Post-publish: tag v0.6.16, GitHub release (not draft, not prerelease), and both npm packages observable at 0.6.16 (@bbingz/polycli-opencode@0.6.16 at 2026-05-20T14:44:18Z, @bbingz/polycli@0.6.16 at 2026-05-20T14:44:33Z).

2026-05-10 — Codex — v0.6.15 observability and provider failure classification

  • Fixed the observability split found after several days of real polycli use: POLYCLI_STATE_ROOT now overrides CLAUDE_PLUGIN_DATA/state, timing --all / --history all can read full history, and timing --json reports store metadata (stateRoot, stateRootSource, workspace slug/root, history limit, record count).
  • Extended timing records with outcome diagnostics (outcome, exitCode, terminationReason, responseMatched, errorCode) so aggregates no longer mix successful health/ask runs with provider failures without explanation.
  • Added run-ledger failure classification and explanation output for failed attempts, including structured counts for qwen max-session-turns, timeouts, terminated/cancelled runs, missing binaries, auth failures, and no-visible-text failures.
  • Hardened provider adapters around observed failure modes: qwen max-session-turns becomes qwen_max_session_turns, kimi resume footer exits with visible assistant text no longer fail, and qwen/kimi/opencode/cmd attach errorCode classifications used by timing and ledger.
  • Updated README/release docs and bumped host/terminal release manifests to 0.6.15; regenerated all five companion bundles.
  • Verification: focused TDD red/green slices for timing, run-ledger, qwen/kimi/opencode/cmd, plus npm test exit 0 (374/374 tests) and npm run release:check exit 0 (full tests, bundle/fixture/manifest/host-map/Codex adapter checks, Claude plugin validation, npm dry-run/pack checks).

2026-05-07 — Codex — v0.6.14 post-publish host update closeout

  • Confirmed npm @bbingz/polycli-opencode@0.6.14 and @bbingz/polycli@0.6.14 are observable on the registry, then updated the GitHub release notes and release docs from "pending npm auth" to published.
  • Verified Codex marketplace refresh behavior: codex plugin marketplace add bbingz/polycli is idempotent and does not refresh an existing local cache; use codex plugin marketplace upgrade polycli-hosts to pull the latest marketplace revision. Local Codex marketplace cache now points at 6e550b3 and polycli-codex@0.6.14.
  • Committed and pushed the post-publish docs closeout as 6e550b3 docs: mark v0.6.14 npm packages published.

2026-05-07 — Codex — v0.6.14 provider path and stateless call hardening

  • Recorded the current best-provider path table in docs/provider-paths.md, including the corrected OpenCode finding: local auth/model discovery is the source of truth, so an empty opencode.json provider object does not mean OpenCode has no configured providers.
  • Hardened prompt/review defaults away from broad YOLO for stateless calls: qwen ask is now bounded at 20 turns with plan mode and tool exclusion instead of the failing one-turn cap; Claude uses no-tools plus empty strict MCP config; Gemini/OpenCode/Pi/Kimi/Cmd/Copilot get conservative provider-specific ask constraints.
  • Kept Copilot as a fallback provider but removed allow-all tool/path/url defaults for ask/review.
  • Replaced MiniMax mini-agent log scraping with official mmx-cli text-chat JSON non-interactive invocation and updated tests/docs/skills around the new path.
  • Verification: npm run check:provider-paths exit 0 with local mmx 1.0.12 included in the drift probe; live polycli ask --provider minimax smoke returned ok: true; release:check exit 0 (367/367 tests; bundles 5; fixtures 16; manifests 0.6.14; host-map 11x4+terminal; codex-adapter 5; claude plugin validate ×2; npm dry-runs/pack checks passed). Post-publish: tag, GitHub release (not draft, not prerelease), and both npm packages observable at 0.6.14.

2026-05-07 — Claude — v0.6.13 released

  • Published artifacts: GitHub release https://github.com/bbingz/polycli/releases/tag/v0.6.13, npm @bbingz/polycli-opencode@0.6.13, npm @bbingz/polycli@0.6.13. Utility packages stay on independent v1.x cadence (@bbingz/polycli-utils@1.0.1, @bbingz/polycli-timing@1.0.1); @bbingz/polycli-runtime remains internal.
  • Default GEMINI_CLI_TRUST_WORKSPACE=true for every gemini spawn. gemini-cli prompts on first run in a new workspace for trust; under polycli's non-interactive ask/rescue/review pipeline that prompt has nowhere to go and the call hangs/fails. New buildGeminiEnv(parentEnv) helper defaults the env var to "true" but preserves any caller-set value (GEMINI_CLI_TRUST_WORKSPACE=false ./script.sh still wins). Consistent with the v0.6.12 YOLO stance: workspace trust is one more interactive prompt that polycli auto-approves.
  • Verification: release:check exit 0 (366/366 tests; bundles 5; fixtures 16; manifests 0.6.13; host-map 11x4+terminal; codex-adapter 5; claude plugin validate ×2; 4 npm pack/publish dry-runs). Post-publish: tag, GitHub release (not draft, not prerelease), and both npm packages observable at 0.6.13.

2026-05-07 — Claude — v0.6.12 released

  • Published artifacts: GitHub release https://github.com/bbingz/polycli/releases/tag/v0.6.12, npm @bbingz/polycli-opencode@0.6.12, npm @bbingz/polycli@0.6.12. Utility packages stay on independent v1.x cadence (@bbingz/polycli-utils@1.0.1, @bbingz/polycli-timing@1.0.1); @bbingz/polycli-runtime remains internal.
  • Standardize ask/rescue permission default to YOLO across all 9 providers. Path B principle: surface real provider capabilities, don't fake uniformity by mixing tiers across providers. Until v0.6.11 polycli's permission stance was an asymmetric mix (copilot/opencode already YOLO; claude acceptEdits; qwen auto-edit; gemini plan; kimi/cmd no flag at all; pi/mini-agent no permission gate). v0.6.12 makes every provider that has a YOLO-equivalent flag pass it by default: claude --permission-mode bypassPermissions, gemini --approval-mode yolo, qwen --approval-mode yolo (also dropped legacy unsafeFlag/background guard), kimi --yolo, cmd --yolo. Callers can opt out via runtime parameters (permissionMode, approvalMode, yolo, skipPermissions). Documented as part of the v1 public surface.
  • review / adversarial-review remain locked to conservative / read-only / plan mode regardless of the new YOLO defaults: review override now adds permissionMode: "plan" for claude, approvalMode: "plan" for qwen (matching the existing gemini override), yolo: false for kimi and cmd. assertNoReviewConstraintOverride extended to refuse downstream callers re-introducing those YOLO flags into a review invocation.
  • Verification: release:check exit 0 (364/364 tests; bundles 5; fixtures 16; manifests 0.6.12; host-map 11x4+terminal; codex-adapter 5; claude plugin validate ×2; 4 npm pack/publish dry-runs). Post-publish: tag, GitHub release (not draft, not prerelease), and both npm packages observable at 0.6.12.

2026-05-07 — Claude — v0.6.11 released

  • Published artifacts: GitHub release https://github.com/bbingz/polycli/releases/tag/v0.6.11, npm @bbingz/polycli-opencode@0.6.11, npm @bbingz/polycli@0.6.11. Utility packages stay on independent v1.x cadence (@bbingz/polycli-utils@1.0.1, @bbingz/polycli-timing@1.0.1); @bbingz/polycli-runtime remains internal.
  • Drop the unilateral 200 KB diff cap on review/adversarial-review: DEFAULT_MAX_DIFF_BYTES flipped from 200_000 to null. With provider context windows now routinely 1M-2M tokens, the hardcoded cap was an artificial cost ceiling that contradicted the Path B "no fake unification" stance. By default the full git diff goes to the provider; callers can still opt into truncation by passing a positive numeric maxDiffBytes to collectReviewContext or --max-diff-bytes <n> on the wrapper.
  • Add --max-diff-bytes <n> CLI flag on review and adversarial-review (validated like --history; invalid_max_diff_bytes structured error code on bad input). Help text, public-surface doc, host-plugin command files (commands/review.md, commands/adversarial-review.md), and codex/copilot SKILL grammar updated to surface the flag.
  • Verification: release:check exit 0 (362/362 tests; bundles 5; fixtures 16; manifests 0.6.11; host-map 11x4+terminal; codex-adapter 5; claude plugin validate ×2; 4 npm pack/publish dry-runs). Post-publish: tag, GitHub release (not draft, not prerelease), and both npm packages observable at 0.6.11.

2026-05-07 — Claude — v0.6.10 released

  • Published artifacts: GitHub release https://github.com/bbingz/polycli/releases/tag/v0.6.10, npm @bbingz/polycli-opencode@0.6.10, npm @bbingz/polycli@0.6.10. Utility packages stay on independent v1.x cadence (@bbingz/polycli-utils@1.0.1, @bbingz/polycli-timing@1.0.1); @bbingz/polycli-runtime remains internal.
  • Pi probe fixes: DEFAULT_PI_MODEL was a hardcoded "openai-codex/gpt-5.4" always injected into the pi command line via buildPiInvocation, breaking probes for any user whose pi was authenticated against a different backend (Xiaomi etc.) — set to null so pi auto-routes to its configured backend. parsePiStreamText now extracts event.message.errorMessage and event.message.stopReason==="error" into a new providerError field; runPiPrompt/runPiPromptStreaming surface that as result.error instead of the generic "pi produced no visible text". event.message.model added to the model extraction paths so reporting reflects the model pi actually used (e.g. mimo-v2.5-pro).
  • Live verification before publish: node packages/polycli-terminal/bin/polycli.mjs health --provider pi --json against a Xiaomi-backed pi flips from ok=false, model="openai-codex/gpt-5.4", error="pi produced no visible text" to ok=true, model="mimo-v2.5-pro", error=null.
  • Verification: release:check exit 0 (359/359 tests; bundles 5; fixtures 16; manifests 0.6.10; host-map 11x4+terminal; codex-adapter 5; claude plugin validate ×2; 4 npm pack/publish dry-runs). Post-publish: tag, GitHub release (not draft, not prerelease), and both npm packages observable at 0.6.10; @bbingz/polycli bin still maps bin/polycli.mjs.

2026-05-07 — Claude — v0.6.9 released

  • Published artifacts: GitHub release https://github.com/bbingz/polycli/releases/tag/v0.6.9, npm @bbingz/polycli-opencode@0.6.9, npm @bbingz/polycli@0.6.9. Utility packages stay on independent v1.x cadence (@bbingz/polycli-utils@1.0.1, @bbingz/polycli-timing@1.0.1); @bbingz/polycli-runtime remains internal.
  • Patch on top of v0.6.8 closing the post-Q6 hardening: dead-worker scan-on-read terminal event recovery in debug runs/show/explain (idempotent attempt_result + provider_decision append; worker_exited classification for no-envelope deaths), TUI log file pointer rendering (read-only — never reads or prints log contents), host-map guardrail extended to Terminal CLI cells / side-by-side examples / terminal-only polycli tui docs, and README command-surface drift cleanup across en/zh/ja so the front matter matches the v0.6.7 + v0.6.8 surface.
  • Verification: release:check exit 0 (354/354 tests; bundles 5; fixtures 16; manifests 0.6.9; host-map 11x4+terminal; codex-adapter 5; claude plugin validate ×2; 4 npm pack/publish dry-runs). Post-publish: tag, GitHub release (not draft, not prerelease), and both npm packages observable at 0.6.9; @bbingz/polycli bin still maps bin/polycli.mjs.

2026-05-07 — Codex — memory and Claude handoff closeout

  • Added project memory for the Q6 terminal CLI/TUI observability track, including v0.6.7/v0.6.8 shipped surfaces, post-v0.6.8 hardening, and the next v0.6.9 release-prep handoff.
  • Added and pushed docs/superpowers/plans/2026-05-07-claude-remaining-work-handoff.md so Claude can fully take over remaining P0/P1/P2 work without reopening completed Q6 implementation scope.

2026-05-07 — Codex — Q6 phase 5-7 hardening closeout

  • Added scan-on-read dead-worker recovery for background runs with residual runContext; debug runs/show/explain refresh job state before reading the ledger and append missing terminal attempt_result / provider_decision events idempotently.
  • Added TUI rendering for local job logFile pointers without reading or printing log contents.
  • Tightened host-map validation so Terminal CLI command cells, side-by-side examples, and terminal-only polycli tui documentation stay in sync.
  • Updated README variants, roadmap, release docs, v0.6.8 notes, and Q6 task state to reflect the shipped @bbingz/polycli terminal package, TUI inspector, and completed post-v0.6.8 hardening.

2026-05-07 — Claude — v0.6.8 released

  • Published artifacts: GitHub release https://github.com/bbingz/polycli/releases/tag/v0.6.8, npm @bbingz/polycli-opencode@0.6.8, npm @bbingz/polycli@0.6.8. Utility packages stay on independent v1.x cadence (@bbingz/polycli-utils@1.0.1, @bbingz/polycli-timing@1.0.1).
  • Includes Q6 Spec 2 — background-job ledger plumbing (runContext persisted into per-job config; parent writes job_started; _job-worker writes attempt_started / attempt_result / provider_decision against the originating runId; worker-observed cancellation produces attempt_result status=cancelled + provider_decision status=cancelled reason=job_cancelled).
  • Includes Q6 Spec 3 — read-only polycli tui inspector (terminal-only; navigation up/down/k/j + enter/b + tab + ? + r refresh + q quit; renders started / attempt_started events without a terminal result as unfinished / unknown; --history <count> validated and applied; raw-mode try/finally with idempotent restore).
  • Real-pty q-exit fix: explicit process.stdin.resume() after setRawMode(true) plus process.stdin.pause() in restoreRawMode(). Previously, in some real-PTY sessions the q keypress never reached the handler and only Ctrl-C escaped (with exit 1).
  • Run-ledger debug examples doc surface (docs/polycli-v1-public-surface.md) walks through the original Q6 narrative (cmd health passed but two ask attempts failed → not adopted; pi health failed → skipped before prompt-bearing work) using polycli debug runs / show / explain.

2026-05-07 — Claude — v0.6.8 release prep (manifests + notes; not published)

  • Bumped 7 release manifests from 0.6.7 to 0.6.8 (9 occurrences): .claude-plugin/marketplace.json, .github/plugin/marketplace.json, plugins/polycli/.claude-plugin/plugin.json, plugins/polycli-codex/.codex-plugin/plugin.json, plugins/polycli-copilot/plugin.json, plugins/polycli-opencode/package.json, packages/polycli-terminal/package.json. Utility packages unchanged on independent v1.x cadence.
  • Added docs/release-notes-v0.6.8.md covering Q6 Spec 2 (background-job ledger plumbing) + Q6 Spec 3 (read-only TUI inspector MVP) + run-ledger debug examples.
  • Updated docs/release.md Current Release State to "prepared for v0.6.8" (last published release still v0.6.7); updated docs/roadmap.md snapshot + Current state to mention prepared-but-not-published status.
  • Recorded automated TUI smoke results in the release notes (script-keys + non-TTY error path against the real wrapper binary). Real-TTY items (q quits + raw-mode restoration, interactive r refresh) explicitly listed as still-needed user-side smoke before tagging.
  • Local npm test 348/348, release:check exit 0 (bundles 5 / fixtures 16 / manifests 0.6.8 / host-map 11×4 / codex-adapter 5 / claude plugin validate ×2 / 4 npm pack dry-runs). No tag, no GitHub release, no npm publish in this slice.

2026-05-07 — Claude — TUI inspector MVP

  • Added terminal-only polycli tui as a read-only inspector over existing debug/run-ledger data.
  • Renders run list, provider states, event timeline, detail/reproduction command panel, and explicit unfinished / unknown states for non-terminal background jobs.
  • New view-model layer (packages/polycli-terminal/lib/tui/view-model.mjs): pure classifyProviderStates / formatReproductionCommand / truncateMiddle / buildTuiModel / renderTuiFrame so behavior is testable without a real TTY.
  • New runtime (packages/polycli-terminal/bin/polycli-tui.mjs): interactive q / r keypress loop over process.stdin raw mode plus a --smoke --fixture-dir <dir> mode for one-frame snapshots used by tests.
  • Terminal wrapper now routes polycli tui to the TUI runtime; all other commands still delegate to the bundled companion. POLYCLI_HOST_SURFACE defaults to terminal for both targets.
  • Packaging: packages/polycli-terminal/package.json now ships bin/polycli-tui.mjs and lib/**/*.mjs; new packaging test asserts bin/polycli-tui.mjs and lib/tui/view-model.mjs are in the published tarball.
  • Docs: terminal README adds a TUI section, host command map adds a terminal-only polycli tui note (no host plugin slash command), public-surface doc adds a polycli tui entry.
  • No provider execution, retry, cancel, daemon, watch mode, full log viewer, version bump, tag, or publish in this slice.

2026-05-07 — Codex — TUI inspector MVP spec/plan

  • Added docs/superpowers/specs/2026-05-07-tui-inspector-mvp-design.md for a read-only terminal TUI inspector over existing debug runs/show/explain and run-ledger data.
  • Added docs/superpowers/plans/2026-05-07-tui-inspector-mvp.md with a task-by-task implementation plan covering view-model tests, terminal runtime, wrapper routing, package files, docs, and release checks.
  • Updated the Q6 roadmap wording to name the read-only TUI inspector MVP as the next implementation slice; no source change, no rebuild, no version bump, no tag, no publish.

2026-05-07 — Claude — docs: run ledger failure examples

  • Added "Run ledger debug examples" section to docs/polycli-v1-public-surface.md, covering the original Q6 narrative: cmd health passed but two ask attempts failed (not adopted); pi health failed and was skipped before any prompt-bearing work. Examples use polycli debug runs / show / explain and reference event-schema slots, not invented live provider output.
  • Closed the "Add docs examples for the concrete failure case" item in tasks/terminal-cli-tui-observability.md.
  • Docs-only; no source change, no rebuild, no version bump, no tag, no publish.

2026-05-07 — Claude — background-job ledger plumbing (Q6 Spec 2)

  • Parent process now persists a top-level runContext (runId / command / hostSurface / argv / jobId / provider / kind / model / defaultModel / logFile) into the per-job config when --run-id (or POLYCLI_RUN_ID) is in scope.
  • After spawning the worker, the parent writes a job_started ledger event; no provider_decision from the parent.
  • _job-worker reads runContext, writes attempt_started before the streaming call, and on completion writes attempt_result (status completed / failed) plus provider_decision (adopted on success, failed reason=<kind>_failed on not-ok). Worker-observed cancellation produces attempt_result status=cancelled + provider_decision status=cancelled reason=job_cancelled.
  • Added shared recordRunEventForContext(workspaceRoot, runContext, base) writer; existing recordRunEvent delegates via buildCurrentRunContext(). Worker code never mutates the parent-side RUN_CONTEXT global.
  • createRunLedgerEvent schema gains nullable pid / durationMs slots; foreground events round-trip with the existing fields and add null defaults for the new ones.
  • Tests: 3 new background integration tests (success with --run-id, failed cmd ask without full prompt leakage, explicit POLYCLI_HOST_SURFACE=codex-skill propagation). All 61 plugin-level tests pass; full npm test and npm run release:check green.
  • Killed-worker (kill -9 after provider returns but before the ledger write) perfect recovery is open ledger-side hardening (reaper or scan-on-read step), not a TUI gate; the first TUI inspector can proceed and must render any started / attempt_started event without a terminal attempt_result / provider_decision as unfinished / unknown.
  • No version bump, no tag, no publish — main only. Roadmap Q6 status updated; tasks/terminal-cli-tui-observability.md background-worker checkbox flipped.

2026-05-07 — Claude — v0.6.7 released (terminal CLI + run ledger)

  • Shipped standalone terminal CLI @bbingz/polycli (PATH-callable wrapper around the bundled companion); first-time npm publish.
  • Added shared debug companion vocabulary (debug runs / show <run-id> / explain <run-id>) surfaced through Claude / Codex / Copilot / OpenCode / terminal.
  • Added redacted append-only run ledger (NDJSON per workspace) with stable runId / workspaceSlug / hostSurface; foreground health, ask, rescue, review, adversarial-review write health_result / attempt_started / attempt_result / provider_decision / run_summary events. Background-worker ledger plumbing is the next follow-up.
  • Added global --run-id <id> (or POLYCLI_RUN_ID) to join multi-command flows under one ledger run; stripped before provider/positional parsing.
  • Bundle / release guards now cover a fifth byte-identical companion bundle (packages/polycli-terminal/bin/polycli-companion.bundle.mjs) and assert the terminal tarball ships LICENSE.
  • Published artifacts: GitHub release https://github.com/bbingz/polycli/releases/tag/v0.6.7, npm @bbingz/polycli-opencode@0.6.7, npm @bbingz/polycli@0.6.7. Utility packages remain on independent v1.x cadence (@bbingz/polycli-utils@1.0.1, @bbingz/polycli-timing@1.0.1).

2026-05-06 — Codex — prepare v0.6.4 Codex install-surface correction

  • Corrected the Codex adapter docs and manifest after real Codex TUI verification showed /polycli-codex:polycli ... is not a registered slash command.
  • Documented the actual Codex flow: codex plugin marketplace add bbingz/polycli, then install Polycli from TUI /plugins, then start a new thread so the polycli skill appears.
  • Reworked the Codex skill invocation guidance to resolve the plugin root from the installed SKILL.md path instead of requiring a manually exported PLUGIN_ROOT.
  • Tightened npm run validate:codex-adapter and npm run validate:host-map so future releases reject fake Codex slash-command examples and require Codex skill examples for health, ask, review, and timing.
  • Prepared release notes and host package metadata for v0.6.4; OpenCode package moves to @bbingz/polycli-opencode@0.6.4, while utils/timing stay on 1.0.1.

2026-05-06 — Codex — prepare v0.6.3 Codex adapter operability patch

  • Strengthened the polycli-codex manifest and skill so Codex is explicitly told to prefer /polycli-codex:polycli ... over direct official CLI shell calls for claude, copilot, opencode, pi, gemini, kimi, qwen, and minimax; raw shell is now documented as an explicit-user-intent or unavailable-plugin fallback only.
  • Updated the root README, Codex plugin README, and host command map with Codex slash examples for health, ask, review, status, result, and timing, so daily examples no longer look like a generic shell command.
  • Added docs/codex-adapter-operability.md as the routing, fallback, first-run, and observability contract for Codex sessions.
  • Added scripts/validate-codex-adapter.mjs, npm run validate:codex-adapter, unit coverage, CI wiring, and release:check wiring so Codex provider triggers, fallback language, and health/status/result/timing guidance cannot drift silently.
  • Prepared release notes and host package metadata for v0.6.3; OpenCode package moves to @bbingz/polycli-opencode@0.6.3, while utils/timing stay on 1.0.1.

2026-05-02 — Claude — opencode added to timeout multiplier + multiway benchmark doc

Triggered by a second 5-round multiway run using real HuggingFace dataset rows (offsets 5/1000/130/400/0 from MMLU college_math / GSM8K / HumanEval/130 / TruthfulQA / BBH), as opposed to the first run which used benchmark-flavored re-creations.

Multiplier extended to opencode:

  • plugins/polycli/scripts/polycli-companion.mjs: PROVIDER_TIMEOUT_MULTIPLIERS.opencode = { "kimi-for-coding/k2p6": 2 }. Same pattern as gemini — model-scoped, only the reasoning variant gets ×2; other opencode models stay at base.
  • Trigger: HumanEval/130 (Tribonacci with awkward forward-reference recurrence) made opencode hit timedOut: true, signal: "SIGTERM" at exactly 120021 ms — the 120s ask ceiling. Verified post-fix that opencode background job timeout is 240000ms (gemini still 240000ms, qwen unchanged at 120000ms).
  • gemini-cli-runtime/SKILL.md Latency expectations section updated to show the multiplier is a registry, not gemini-only — the same pattern can be extended as new reasoning model ids appear.

Benchmark doc persisted:

  • docs/benchmarks/multiway-validation-2026-05-02.md captures: exact dataset offsets and prompt verbatim text, full result matrix for both rounds (40/40 easy + 37/40 hard), per-provider behavior notes (kimi misreads complex prompts, minimax reasons shallowly on puzzle-class, opencode is a code-reasoning model), and 3 grader bugs that bit me during the run (.strip() ate Python indent, prefix-only negation regex missed kimi's "not because" answer, single-language regex misclassified claude's Chinese reply). The grader-lessons section is the durable artifact — future benchmark scripts should not repeat these.

What we did NOT do:

  • Did not write a reusable benchmark grader utility module (tasks/benchmark-grader.mjs style). polycli is not a benchmark suite; persisting the lessons in a doc is sufficient.
  • Did not extend the multiplier to other providers preemptively — opencode addition was driven by a real timeout observation, not speculation.

2026-05-02 — Claude — kimi/minimax model id fix + gemini multiplier stress test

Followups discovered during a 5-round 9-way validation (8 polycli providers + codex via subagent, prompts from MMLU / GSM8K / HumanEval / TruthfulQA / open-ended).

Fix: kimi & minimax model field was null in ask result

  • packages/polycli-runtime/src/kimi.js:347 and :404 — fall back to readKimiDefaultModel() (already existed but was unused) when parsed.model ?? model ?? defaultModel is all null. Reads ~/.kimi/config.toml default_model scalar.
  • packages/polycli-runtime/src/minimax.js:238 — fall back to readMiniMaxConfig().model (reads ~/.mini-agent/config/config.yaml) under the same condition.
  • Root cause: cacheProviderModel only writes when result.model is non-empty. kimi stream-json and mini-agent log don't carry a model id, so cache stayed empty, so subsequent runs got null defaultModel, so model stayed null. Chicken-and-egg. Direct config read breaks the loop.
  • Verified post-fix: kimi returns kimi-code/kimi-for-coding, minimax returns MiniMax-M2.7-highspeed. Resolves the "null for kimi/mini-agent" caveat in reference_default_model_extraction_caveats.md.

Stress test: gemini timeout multiplier

Goal: prove the gemini-3.1-pro-preview ×2 multiplier was actually necessary, not just defensive.

prompt class gemini wall gemini ttft observation
Standard 5-round (MMLU / GSM8K / HumanEval / TruthfulQA / open-ended) 7-22s (mostly < total) far below 120s base — multiplier irrelevant
GPQA-style physics reasoning + algebra (453 byte prompt, ask) 55s 44s half the base 120s — comfortable headroom but base would have worked
Heavy structured output: 800-byte rate-limit design prompt (rescue) 55s 37s + 18s gen far below 600s rescue base

Result: under prompts I could construct, gemini-3.1-pro-preview peaks around 55s — never approaches even the 120s ask base, much less the 240s multiplier ceiling. The multiplier is defensive headroom for the original observation (user reported gemini "self-admitted long thinking time" on review of large diffs), not a tight fit. Documented this honestly: the multiplier provides safety margin for the worst case rather than reacting to a routine ceiling breach. No timeout tuning change.

Side observations from the 5-round validation (no fixes needed, recorded for posterity):

  • All 9 entities (8 polycli + codex) returned correct answers on R1/R2/R4. Code-completion (R3) instruction-following varied: 7/8 polycli wrapped the function in markdown despite "no extra commentary" — only minimax and codex gave a true one-liner. This is upstream LLM behavior, not polycli routing.
  • claude provider answered the English R4 prompt in Chinese because the user's global CLAUDE.md says "Always respond in Chinese". CLAUDE.md inheritance only reaches the claude provider; other providers respect prompt language. Documented in memory as expected behavior.
  • 5 rounds × 8 providers wall time: 113s (parallel start per round, sequential rounds). Codex reference run (subagent a48de540074220012): all 5 prompts correct, 7-9s latency each.

2026-05-02 — Claude — multi-way self-review: 4 doc fixes + 1 verified bug + multiplier scope tightening

Ran a 4-way self-review where each provider audited polycli's claims about itself (gemini / qwen / kimi / minimax each reading their own *-cli-runtime/SKILL.md). Triaged findings into red (true defects, fix), yellow (LLM claims to verify against real CLIs), and green (engineering choices).

Red — 4 doc defects fixed:

  • minimax-cli-runtime/SKILL.md P0.5: removed self-contradiction. Section claimed Layer 1/3 sentinels were "跨 locale 稳定" while the same paragraph noted OSError messages may be i18n'd by glibc. Reworded as "纯 ASCII 字面量, 未观察到 i18n" and clarified the i18n caveat is OS-layer (outside Mini-Agent's control).
  • minimax-cli-runtime/SKILL.md P0.9: scoped the "0 次 os.environ" claim to first-party Mini-Agent code (excludes transitive deps like httpx / pydantic) and clarified the implication is auth-purpose only.
  • qwen-cli-runtime/SKILL.md: documented --unsafe vs --approval-mode precedence with source-verified semantics. Initial draft said "--unsafe wins", but reading buildQwenInvocation in packages/polycli-runtime/src/qwen.js:73-76 showed the opposite: an explicit --approval-mode wins; --unsafe is only a shortcut to yolo when --approval-mode is omitted. Background-mode yolo still requires --unsafe as an independent safety guard.
  • kimi-cli-runtime/SKILL.md: "Auth ping" was misleading — --max-steps-per-turn 1 + 30s is a liveness probe (verifies binary launches and reaches the model) but does not validate token freshness. Renamed to "Liveness probe" and pointed to setup --json authenticated field for true auth state.

Yellow — 6 claims verified against real CLIs (gemini 0.40.1, kimi 1.40.0, qwen 0.15.6):

# Claim Verdict
5 gemini --write → --approval-mode auto_edit ✅ help confirms auto_edit is a valid enum
6 kimi -V and -v flags ⚠️ partially wrong — -V and --version both work, but -v returns a click usage error in 1.40.0+, not verbose. Doc updated.
7 kimi stream-json role set ✅ live probe shows only assistant (kimi's self-claim of "user/system" was a hallucination)
8 kimi --approval-mode acceptance ✅ kimi 1.40 has no --approval-mode (only --yolo/--plan/--afk/--print); doc was correct
9 --print ≈ --yolo ✅ kimi --help 1.40 says "Print mode auto-dismisses AskUserQuestion and auto-approves tool calls"; doc was correct
10 kimi auth paths ✅ subcommand list is login/logout/term/acp/info/export/mcp/plugin/vis/web — no API-key/SSO subcommands; doc was correct

Net: kimi's self-review hallucinated 5 of 6 — LLMs are unreliable narrators about their own CLI surfaces. Real probes are necessary. Only Q6 produced a doc edit (-v wording in kimi-cli-runtime/SKILL.md lines 36 and 97).

Green — multiplier tightened to model scope:

  • Refactored PROVIDER_TIMEOUT_MULTIPLIERS = { gemini: 2 } → { gemini: { "gemini-3.1-pro-preview": 2 } }. Refactored resolveTimeoutMs signature to (provider, kind, { model, defaultModel }). Resolution: explicit --model wins; falls back to cached upstream-default model only if caller did not pass one. So --model gemini-flash-2.5 stays at base 120s/600s/300s even though the cached default is a reasoning model — addresses gemini self-review's complaint that the multiplier was over-broad.
  • gemini-cli-runtime/SKILL.md: Latency expectations rewritten with explicit resolution rules and a note that adding new reasoning model ids (when upstream releases them) is the maintenance path — do not blanket-multiply the whole provider.
  • qwen-cli-runtime/SKILL.md: Safety rule "If Bash call fails, return nothing" gets a Rationale clarifying this is a forwarder contract — the companion already encodes failure in error/exit code, the subagent re-emitting prose would only duplicate or paraphrase. (Addresses qwen self-review #2.)

Side check — mini-agent upstream: verified via GitHub API (/repos/MiniMax-AI/Mini-Agent): no releases, no tags, latest commit d76a4f63 2026-02-14 (cosmetic fix). Still 0.1.0; minimax self-review's hint about "possibly newer version" was unsupported. Memory entry stays accurate.

2026-05-02 — Claude — README clarifies polycli is an in-host plugin, not a shell binary

  • Triggered by a real Codex session that grepped the user's PATH directories for a polycli binary, found nothing, concluded "polycli has no callable entry point", and fell back to invoking qwen directly — defeating the routing purpose of polycli. Codex-rescue review (agent afee5f7a594387551) confirmed the misleading signals: hero SVG showed $ polycli health --json (shell prompt), and README L25/L36/L108 used cross-host vocabulary phrasing that read as if polycli were a portable shell command.
  • docs/assets/readme-header.svg:54: $ polycli health --json → /polycli:health --json. Removes the shell-prompt visual cue from the first thing every reader sees.
  • README.md "What is polycli?" gets a callout: polycli is an in-host plugin, no polycli binary on PATH, each host adapter exposes the same vocabulary in its own invocation style. Quick-start Copilot row annotated as "skill word — NOT a PATH binary; only inside the copilot prompt" so that line cannot be screenshot-grepped out of context.
  • README.md adds ## Outside a supported host section listing three honest options for non-Claude-Code/Codex/Copilot/OpenCode agents: (1) install the host adapter for the environment, (2) call the underlying provider CLI directly with explicit trade-offs, (3) escape hatch — PLUGIN_ROOT=... node scripts/polycli-companion.bundle.mjs ... marked unstable and internal. Verified against plugins/polycli-codex/skills/polycli/SKILL.md:11 (uses ${PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-}}).
  • README.zh-CN.md and README.ja.md get the in-host callout + Quick-start row clarification, but link back to the English Outside a supported host section instead of duplicating the full block (translation drift > terse cross-link).
  • Docs only — no code, no test changes.

2026-05-02 — Claude — gemini timeout multiplier + latency documentation

  • plugins/polycli/scripts/polycli-companion.mjs introduces PROVIDER_TIMEOUT_MULTIPLIERS = { gemini: 2 } and a resolveTimeoutMs(provider, kind) helper. Applied to the two prompt-execution code paths (parsePromptExecution for ask/rescue, runReview for review/adversarial-review). health retains the universal 60s budget — gemini's health probe is sentinel-only and does not exercise reasoning.
  • Effective ceilings for gemini: ask 240s (was 120s), rescue 1200s (was 600s), review / adversarial-review 600s (was 300s). All other providers unchanged.
  • Motivation: gemini is a deep-reasoning model that routinely spends 30s–several minutes thinking before emitting visible text (live observation across 2026-04-29 bench + this session: rescue PONG took 44.7s vs qwen 2.9s on identical prompt). Hard-coded 300s review ceiling was the closest one to silently masquerade as "polycli broken".
  • plugins/polycli/skills/gemini-cli-runtime/SKILL.md adds a ## Latency expectations section explaining the multiplier, listing the new vs old ceilings per kind, and recommending --background + /polycli:status polling for prompts of unknown duration. Subcommand table updated to reflect the gemini-specific timeouts.
  • Background worker (runJobWorker) reuses execution.timeout, so the multiplier propagates to background jobs as well — no separate code path needed.

2026-05-02 — Claude — fix stale task subcommand references in 4 cli-runtime SKILLs

  • plugins/polycli/skills/{qwen,gemini,kimi,minimax}-cli-runtime/SKILL.md referenced a task companion subcommand that does not exist on the unified surface (companion exposes setup, health, ask, rescue, review, adversarial-review, status, result, cancel, timing). The references were inherited verbatim from the legacy *-plugin-cc repos in R8c (commit 193078f) where each plugin had its own companion with a task command; on the unified surface that role split into ask (120s, one-shot) and rescue (600s, multi-step). A polycli:polycli-provider-agent subagent that read the SKILL literally would have invoked task and crashed with Unknown subcommand 'task'.
  • Same edit also dropped non-existent task-resume-candidate (resumable state lives behind --resume-last) and the legacy write-key helper from minimax (no longer on the unified surface).
  • End-to-end smoke: polycli-companion.bundle.mjs rescue --provider qwen returns ok=true, kind=rescue, response present (real qwen-cli 0.15.6, qwen3.6-plus). ask --provider qwen already worked.
  • Docs only — no companion / runtime / bundle changes.

2026-04-30 — Claude — landscape check + daily watcher for official LLM-provider plugins

  • Verified openai/codex-plugin-cc (the only LLM-provider-official Claude Code plugin) is at v1.0.4 / 807e03a and matches the locally installed marketplace clone — no upstream updates since 2026-04-18.
  • Surveyed 25+ AI-provider GitHub orgs, the *-plugin-cc naming convention, and Anthropic's claude-plugins-official (171) + claude-plugins-community (1921) marketplaces. Result: as of 2026-04-30, OpenAI is still the only LLM provider with an official Claude Code plugin. Google / xAI / Mistral / DeepSeek / Qwen / Kimi / MiniMax / Cohere / Meta / Groq / Zhipu / 01-ai / Perplexity / OpenRouter — all absent. Microsoft has platform skills (skills-for-fabric, power-platform-skills, skills-for-copilot-studio) but no LLM-provider bridge.
  • Created daily-running remote routine trig_01RLU5aqzYkuPFA8LMQKcnzo (https://claude.ai/code/routines/trig_01RLU5aqzYkuPFA8LMQKcnzo) to watch this signal: 02:00 UTC every day, sonnet-4-6, scans the same orgs/marketplaces and is loud only when a new official provider plugin appears.
  • No code changes in polycli itself.

2026-04-29 — Claude — capability matrix for workflows with no bare-shell equivalent

  • Added docs/benchmarks/capability-matrix.md listing workflows where bare-shell has no meaningful equivalent: adversarial-review, background job control, session resume, stop-review-gate hook, 4-state timing, multi-host consistency (Claude Code / Codex / Copilot / OpenCode), provider health probe, probing-cost amortization. Companion to bench-vs-bare-cli-spec.md. These are presence/absence claims, not byte ratios — forcing them into a token comparison would be dishonest because there's nothing to compare against.
  • README "Cost vs raw shell calls" section gets a one-line pointer to the capability matrix.
  • Spec followup #3 (Path b disciplined-bare-shell) closed as permanent deferral: the (a) vs (c) data already gives a clear directional answer (69–98% reduction across cells, advantage from probing-cost amortization). Cost-benefit (Anthropic SDK dep + ~18 paid API calls per re-run) does not justify the refinement.

2026-04-29 — Claude — close CLAUDE.md inheritance investigation + small bench doc fixes

  • Investigated the CLAUDE.md inheritance question raised in pilot (parent answered in Chinese, polycli claude in English on the same review task). Not a bug. CLAUDE.md inheritance works at the CLI layer (claude -p) and the polycli ask command — both respect the user's language preference. The English output observed in pilot came from buildReviewPrompt (plugins/polycli/scripts/lib/review.mjs), whose template is hardcoded English ("You are acting as inside polycli.", etc.); combined with the English diff payload, models default to English for review/rescue/adversarial-review. Treated as feature, not bug — multi-provider review benefits from a consistent prompt baseline. Spec followup updated.
  • README: clarified "Bare-shell + probing" column meaning with a footnote pointing to probing-cost.json. The previous header could read as raw response bytes; clarification follows Codex review feedback.
  • docs/benchmarks/results-2026-04-29.md: added a caveat noting that qwen review run #2 returned 1 byte at exitCode: 0 (treated as success because exit code, not body, gates the bench's failure count). Median is unaffected; flagged for transparency.

2026-04-29 — Claude — add claude-prompting skill + bench path sanitization

  • Added plugins/polycli/skills/claude-prompting/SKILL.md. Polycli previously had gemini-prompting / kimi-prompting / qwen-prompting / minimax-prompting but no per-provider prompt scaffolding for the claude provider. The new skill encodes claude-specific prompting guidance (CLAUDE.md does not propagate, same model family, stateless by default, error-surface notes).
  • Hardened scripts/bench-vs-bare-cli.mjs with path sanitization: replaces $HOME with ~ in all stored stdout/stderr/parsedJson before writing results JSON. Earlier run leaked maintainer-local paths into results-2026-04-29.json because qwen rescue output hallucinated absolute paths under the bench cwd; scripts/tests/open-source-hygiene.test.mjs caught it.
  • Re-sanitized the published docs/benchmarks/results-2026-04-29.json retroactively. npm test now passes 287/287.

2026-04-29 — Claude — first benchmark of polycli vs bare-shell CLI invocation

  • Added scripts/bench-vs-bare-cli.mjs: N=3 live bench comparing path (a) Bash(<provider> -p) and path (c) polycli-companion for gemini/qwen × ask/review/rescue. Outputs docs/benchmarks/results-<date>.{json,md}.
  • Collected docs/benchmarks/probing-cost.json (lower bound: which + --help only): gemini 3843 B, qwen 8077 B.
  • Headline: with probing cost amortized, polycli reduces parent-context bytes by 69-98% across all six scenario × provider cells. Without probing cost, boundary bytes vary by cell with no consistent direction. polycli's token advantage comes from invocation-knowledge encapsulation, not output compression.
  • Added README "Cost vs raw shell calls" section pointing at the results.
  • Spec lifecycle (tasks/bench-vs-bare-cli-spec.md): two rounds of Codex sign-off, post-pilot amendment switching from fixture replay to live CLI calls (fixture replay was erasing probing cost), and a final Codex round-3 review that surfaced three blocks (missing rescue scenario, raw stdout not preserved for diagnosis, ±15% noise claim incorrect) — all fixed before publish.
  • Path (b) disciplined-bare-shell deferred — needs Anthropic SDK to drive Claude programmatically; pilot data suggests (b) and (c) are close on boundary bytes.
  • Followups noted in spec: add claude-prompting skill (no per-provider scaffolding for the claude provider in polycli today); investigate global CLAUDE.md inheritance into polycli subagent.

2026-04-29 — Codex — close post-release maintenance

  • Merged all open Dependabot PRs after the v0.6.2 publication: actions/setup-node 4 -> 6, actions/checkout 4 -> 6, and zod 4.1.8 -> 4.3.6.
  • Confirmed the public repo has no open PRs, main is clean at fe4c6d6, the latest release remains v0.6.2, and the published npm packages remain aligned with the release notes.
  • Re-ran npm test, npm run release:check, and npm audit --audit-level=moderate; all passed, with 287/287 tests and 0 vulnerabilities.
  • Confirmed GitHub Actions CI succeeded on the three post-release main push runs. GitHub social preview remains a repository settings UI upload using docs/assets/social-preview.png; GitHub CLI exposes no social preview image option.

2026-04-29 — Codex — finish v0.6.2 public release polish

  • Added GitHub Actions CI, README release/OpenCode badges, and a social preview PNG derived from the README header SVG.
  • Tightened public package metadata for @bbingz/polycli-opencode, @bbingz/polycli-utils, and @bbingz/polycli-timing.
  • Replaced the long release:check shell command with scripts/check-release.mjs.
  • Expanded open-source hygiene scanning to all tracked files and archived historical review/session docs under docs/archive/.

2026-04-29 — Codex — prepare v0.6.2 open-source hygiene patch

  • Prepared host plugin manifests and OpenCode package for 0.6.2; prepared @bbingz/polycli-utils and @bbingz/polycli-timing for 1.0.1.
  • Fixed the timing package tarball so @bbingz/polycli-timing/schema resolves to a packed timing.schema.json.
  • Added package-local MIT LICENSE files for all public npm packages and open-source packaging tests that verify export targets and license inclusion from real npm pack --dry-run --json output.
  • Removed the OpenCode adapter dependency on @opencode-ai/plugin, replacing it with a tiny local wrapper plus zod; npm audit --audit-level=moderate now reports 0 vulnerabilities.
  • Scrubbed public fixtures and package AGENTS files of maintainer-local paths, host auth metadata, local memory metadata, and provider reasoning signatures; added a hygiene regression test for those patterns.
  • Replaced the flaky wall-clock concurrency assertion in the health integration test with fake-provider start/end event overlap verification.
  • npm run release:check passes end-to-end with 286/286 tests and publish dry-runs for @bbingz/polycli-opencode@0.6.2, @bbingz/polycli-utils@1.0.1, and @bbingz/polycli-timing@1.0.1.

2026-04-29 — Claude — v0.6.1 docs/legal patch shipped

  • Bumped 6 manifest/package versions from 0.6.0 to 0.6.1: plugins/polycli-opencode/package.json, .claude-plugin/marketplace.json (× 2 entries), .github/plugin/marketplace.json (× 2 entries), plugins/polycli-codex/.codex-plugin/plugin.json, plugins/polycli/.claude-plugin/plugin.json, plugins/polycli-copilot/plugin.json. @bbingz/polycli-utils and @bbingz/polycli-timing stay at 1.0.0 (no source changes).
  • Drafted docs/release-notes-v0.6.1.md — docs/legal patch only: README rewrite + i18n (en/zh-CN/ja), root LICENSE, dead-absolute-path fix, latent fix specs filed in tasks/.
  • Updated docs/release.md "Current Release State" to v0.6.1.

2026-04-29 — Claude — README rewrite + i18n + LICENSE

  • Rewrote README.md from scratch as international-standard, English-default. Added clear hero pitch, "Why polycli" differentiation (4-state timing honesty, no fake unification, direct CLI passthrough), badges (npm version × 2, MIT license, Node ≥20), and a language switcher.
  • Added translations: README.zh-CN.md (Simplified Chinese) and README.ja.md (Japanese). All three are full peers — not abbreviated versions. Technical terms (runtime, streaming, session resume, Path B, monorepo) kept in English by convention; Japanese version uses です・ます style.
  • Added root LICENSE file (MIT, Copyright (c) 2025 bbingz) — the sub-packages already declared MIT, the root file was missing. GitHub licenseInfo was previously null; this fixes the License badge target and makes the project legally complete by community standards.
  • Fixed dead links: previous README contained absolute paths like /home/user/-Code-/polycli/... that did not work on GitHub. All internal links are now repo-relative. Verified every referenced path exists.
  • Restructured: hero → why → hosts/providers → install → quick start → core commands → capability matrix → timing semantics → packages → development → release → contributing → license. Old structure (Who This Is For, Background Jobs, Current Scope) removed or merged.

2026-04-29 — Claude — provider CLI upgrades + default-model audit

  • Upgraded local provider CLIs to upstream latest: copilot 1.0.35 → 1.0.39, kimi-cli 1.37.0 → 1.40.0, mini-agent (git+main) refreshed to 2026-02-14 commit (deps: pydantic 2.13.2 → 2.13.3, uvicorn 0.44 → 0.46, sse-starlette 3.3 → 3.4). Five other CLIs (claude, gemini, qwen, opencode, pi) already at upstream latest.
  • Verified kimi 1.40 and refreshed mini-agent stream-json compatibility via live probes. kimi.test.js 13/13 still pass; live runKimiPromptStreaming extracts response and sessionId correctly. mini-agent live probe parses 48 progress events, strips ANSI, extracts response.
  • Ran 8-CLI default-model audit by spawning each provider with a "what model are you" prompt and reading polycli's result.model. Surfaced two latent bugs:
    • gemini.js:135 takes Object.keys(parsed.stats?.models ?? {})[0] — first attempted model, not actually-used. Misleads when gemini-cli auto-falls-back from a 429-throttled preview (e.g. gemini-3.1-pro-preview → gemini-2.5-pro due to Google server-side preview capacity).
    • kimi.js:174/:264 has readKimiDefaultModel() reading ~/.kimi/config.toml, but it's only consumed by getKimiAuthStatus, never threaded into runKimiPromptStreaming / runKimiPrompt results — so result.model is null even when config has a default.
  • Both fixes drafted as specs in tasks/model-extraction-fixes.md for Codex implementation. Non-breaking; target v0.6.x patch or v0.7.
  • Memory: added reference_cli_provider_versions.md (per-CLI version-check + upgrade commands + gotchas), reference_default_model_extraction_caveats.md (which providers' result.model is unreliable and why), and feedback_no_ask_for_nondestructive.md (skip confirmation gates for sandboxed/read-only ops).

2026-04-24 — Claude — close R8 bookkeeping post-ship

  • Relaxed the CLAUDE.md legacy-repo constraint to a permanent "allow grep, no edits" form (dropped the "R8 期间" conditional now that R8 is complete and R8g was cancelled).
  • Updated project_legacy_repos_reference.md memory from "R8 convergence targets" to "permanent references; v0.6.0 absorbed their functionality".
  • Added two new feedback memory entries: codex-rescue operational quirks (fire-and-forget wrapper, sandbox git-commit block, branch-name misreport) and release-ops gotchas (claude plugin validate Node 25 crash, npm 2FA TTY requirement).

2026-04-24 — Claude — v0.6.0 shipped

  • Pushed main and tag v0.6.0 (tag at a95e3d8).
  • GitHub release live at https://github.com/bbingz/polycli/releases/tag/v0.6.0 with notes from docs/release-notes-v0.6.0.md; no tarball asset attached (OpenCode users install from npm).
  • Published @bbingz/polycli-opencode@0.6.0 to npm; confirmed via npm view @bbingz/polycli-opencode versions.
  • Closes roadmap R8a-R8f. R8g (legacy repo archival) deliberately skipped per user direction ("不用 archive"); four legacy plugin repos remain as read-only references on GitHub without archival status. @bbingz/polycli-utils@1.0.0 and @bbingz/polycli-timing@1.0.0 unchanged from v0.5.0; runtime stays private.

2026-04-24 — Claude — draft v0.6.0 release notes for R8 convergence

  • Added docs/release-notes-v0.6.0.md covering R8a-R8e deliverables with a full legacy → polycli migration table (slash commands, kimi session flags, gemini approval/effort flags, subagent types, guidance skills, hooks).
  • Flagged R8f as done; R8g (legacy repo archival + CLAUDE.md relaxation) still pending.
  • Status: draft. Release date and version-bump execution held until user kicks off release prep.

2026-04-24 — Codex — port stop-time review gate hooks

  • Added Claude Code SessionStart / SessionEnd / Stop hook registration for the polycli host plugin, with lifecycle state cleanup and the optional stop-review gate.
  • Recorded the last-used ask/rescue provider in workspace state so the gate honors the user's current provider selection, with health-probe fallback when no provider is recorded.
  • Added hook tests for lifecycle cleanup, multi-line ALLOW/BLOCK sentinel parsing, timeout skip behavior, and unresolvable-provider skip behavior.

2026-04-24 — Codex — add R8a/R8d unified flags

  • Added Kimi-only --resume-last / --resume <uuid> / --fresh handling on /polycli:ask and /polycli:rescue, including wrapper-side session validation and resume-mismatch warnings.
  • Added Gemini-only --write and --effort low|medium|high handling on the unified ask/rescue surface.
  • Documented and tested unsupported-provider silent-drop notes for the new provider-specific flags.

2026-04-24 — Claude — open R8 legacy plugin convergence

  • Reversed the prior non-goal "No migration of legacy sibling repos into this monorepo" after a capability gap audit against gemini-plugin-cc / kimi-plugin-cc / qwen-plugin-cc / minimax-plugin-cc.
  • Added R8 to docs/roadmap.md with sub-items R8a–R8g: kimi session continuation; session-lifecycle + stop-time review gate hooks; per-provider guidance skills; rescue flag semantics; per-provider subagent types; namespace UX; legacy repo retirement.
  • Updated project_legacy_repos_reference.md memory to flip the stance from "reference-only, never migrate" to "convergence targets under R8; grep-for-port OK, no edits."
  • Narrowed the CLAUDE.md architecture-boundary constraint on legacy repos from "不要 grep、不要编辑" to "不要编辑; R8 期间允许 grep-for-port" (user-confirmed in same session).

2026-04-24 — Codex — v0.5.1 shipped

  • Pushed main and tag v0.5.1 (tag at 0b79c86).
  • GitHub release live at https://github.com/bbingz/polycli/releases/tag/v0.5.1 with notes from docs/release-notes-v0.5.1.md; no tarball asset attached because OpenCode users install from npm.
  • Published @bbingz/polycli-opencode@0.5.1 to npm and confirmed it appears in npm view @bbingz/polycli-opencode versions.

2026-04-24 — Codex — prepare v0.5.1 release

  • Bumped the four host plugin manifests and Claude/Copilot marketplace metadata from 0.5.0 to 0.5.1.
  • Drafted docs/release-notes-v0.5.1.md for the Q2/Q3 guardrail patch release.
  • Release scope is host/plugin line only: publish @bbingz/polycli-opencode@0.5.1; keep utils/timing at 1.0.0 and runtime private.

2026-04-24 — Codex — close Q2 and Q3 with guardrails

  • Closed Q2 by documenting the model fallback policy and adding a host integration test that proves cached setup model metadata is used only as the final defaultModel fallback when a provider stream omits model fields.
  • Closed Q3 by accepting host-surface asymmetry as the durable design and adding npm run validate:host-map to keep docs/host-command-map.md, Claude commands, Codex/Copilot skills, OpenCode tools, and the companion dispatcher aligned.
  • Refreshed docs/roadmap.md to mark R1-R7 and Q1-Q3 closed after v0.5.0.

2026-04-24 — Claude — v0.5.0 shipped

  • Pushed main + tag v0.5.0 (tag at 306c703 — R5 fixture replay pilot commit).
  • GitHub release live at https://github.com/bbingz/polycli/releases/tag/v0.5.0 (notes from docs/release-notes-v0.5.0.md).
  • npm: first public publishes of @bbingz/polycli-utils@1.0.0 and @bbingz/polycli-timing@1.0.0; @bbingz/polycli-opencode@0.5.0 also published. All three confirmed via npm search '@bbingz/'.
  • Closes roadmap R5 (Claude host fixture pilot), R6 (auth-probe transient-error regex named contracts), and Q1 (utils + timing published). Q2 and Q3 remain in "observing" state. Runtime stays private. 256/256 tests pass at release tag.

2026-04-24 — Claude — prepare v0.5.0 release

  • Bumped the four host plugin release manifests and both host marketplace metadata entries from 0.4.2 to 0.5.0.
  • Drafted docs/release-notes-v0.5.0.md covering R6 (auth-probe regex contracts), P1/Q1 (utils + timing first public npm publish), and R5 (Claude fixture replay pilot).
  • release:check passes end-to-end: 256/256 tests, manifest validation, plugin validation, @bbingz/polycli-opencode@0.5.0 / @bbingz/polycli-utils@1.0.0 / @bbingz/polycli-timing@1.0.0 dry-run publishes all succeed.

2026-04-24 — Codex — pilot claude host integration fixture replay

  • Added captured real Claude CLI stream fixtures for host-level /ask and /health success coverage.
  • Added a host integration replay helper that surfaces missing fixture files explicitly and replays captured stream output through the bundled companion without changing other provider tests.
  • Converted the Claude /ask success path to fixture replay and added a Claude /health success integration test that records timing from the replayed stream.

2026-04-24 — Codex — prepare utils and timing for first npm publish

  • Made @bbingz/polycli-utils and @bbingz/polycli-timing publishable by removing private: true and adding npm metadata, package file lists, and public publish config while leaving @bbingz/polycli-runtime private.
  • Refreshed both package READMEs and rewrote docs/polycli-v1-public-surface.md as the live v1 contract for utils + timing only.
  • Added pack:utils / pack:timing scripts, updated manifest validation to require only runtime to stay private, and made release:check validate publishable packages repeatably: unpublished versions use npm publish --dry-run, already-published versions fall back to npm pack --dry-run.

2026-04-24 — Codex — name auth-probe transient-error regex contracts

  • Extracted the transient auth-probe regexes in gemini / qwen / kimi / opencode / pi into provider-local TRANSIENT_PROBE_ERROR_PATTERNS exports without changing the existing match behavior.
  • Added provider tests that assert the named patterns keep timeout-like probe failures in the loggedIn=true / inconclusive path while still treating explicit 401 Unauthorized errors as genuine auth failures.

2026-04-24 — Claude — decide Q1 (publish utils/timing); annotate Q2/Q3 as observing

  • User directive on 2026-04-24: publish @bbingz/polycli-utils and @bbingz/polycli-timing to npm ("能用就应该发"). Extended docs/archive/review-2026-04-24-v0.5.0-spec.md with a P1 (Q1) section covering manifest fields, README refresh, v1-public-surface rewrite, pack:utils / pack:timing scripts, and release:check extension with dry-run publishes.
  • Q2 (model fallback sustainability) and Q3 (four-host surface convergence) marked "observing" in docs/roadmap.md per user direction: record the concern, do not act until a signal accumulates.
  • Runtime (@bbingz/polycli-runtime) stays "private": true; this decision applies only to utils + timing.

2026-04-24 — Claude — close roadmap R4 + R7; refresh roadmap for v0.5.0

  • Added docs/host-command-map.md — capability × host mapping for all 10 polycli commands across Claude Code / Codex / Copilot / OpenCode plus side-by-side invocation examples. Closes roadmap R4.
  • Added scripts/check-review-cli-drift.mjs that probes installed provider CLIs and asserts the flags /review hard constraints depend on (--tools / --approval-mode / --policy / --excluded-tools / --agent / --no-tools) still exist. Env-var based constraints (MiniMax MINI_AGENT_CONFIG_PATH, OpenCode OPENCODE_CONFIG_CONTENT) are listed as manual-watch reminders. Wired as npm run check:review-drift. Closes roadmap R7.
  • Refreshed docs/roadmap.md: removed the completed R1 / R2 / R3 / R4 / R7 entries, updated Current state to v0.4.2 / 250 tests, narrowed v0.5.0 scope to R5 + R6 + optional Q1/Q2/Q3 decisions.

2026-04-24 — Claude — v0.4.2 shipped

  • Pushed main + tag v0.4.2 (tag at 5c7f709 — R2 runtime fix commit).
  • GitHub release live at https://github.com/bbingz/polycli/releases/tag/v0.4.2 (notes from docs/release-notes-v0.4.2.md).
  • npm: @bbingz/polycli-opencode@0.4.2 published.
  • Closes roadmap R1 (p2-p3-backlog merged; 511fceb dropped as superseded by 12d9ca9), R2 (minimax capability matrix + review.mjs YAML hardening + tmp cleanup), and R3 (v1 surface doc superseded). 250/250 tests pass at release tag.

2026-04-24 — Claude — prepare v0.4.2 release

  • Bumped the four host plugin release manifests and both host marketplace metadata entries from 0.4.1 to 0.4.2.
  • Drafted docs/release-notes-v0.4.2.md covering the R1 (spawn cancellation / arg parsing / stream JSON scalars / performance.now timing) and R2 (review.mjs YAML scalar + tmp cleanup) roadmap closures.
  • release:check passes end-to-end: 250/250 tests, manifest validation, plugin validation, @bbingz/polycli-opencode@0.4.2 dry-run publish.

2026-04-24 — Codex — harden review config parsing and temp cleanup

  • Replaced MiniMax review config scalar extraction with a small private parser that supports plain, single-quoted, and double-quoted scalars while rejecting unsupported block/folded and malformed YAML forms.
  • Registered review temp files for best-effort cleanup on process exit so generated per-review config directories do not accumulate across long-running hosts.
  • Added regression coverage for supported scalar forms, unsupported block scalars, malformed config lines, comment handling, and child-process exit cleanup.

2026-04-24 — Codex — land P2/P3 runtime backlog hardening

  • Landed the surviving p2-p3-backlog runtime commits as one squash: spawnStreamingCommand now supports AbortSignal cancellation, stdout decoder overflow protection, stdin drain handling, and post-settle output suppression.
  • Hardened runtime reliability around MiniMax log-read failures, provider exit error formatting, and registry timing by switching prompt duration measurement to monotonic performance.now().
  • Tightened shared utility and timing edge cases: argument parsing rejects malformed empty booleans / unterminated quotes, stream JSON parsing recognizes arrays and bare scalar values, and timing validation rejects invalid numeric bounds.

2026-04-24 — Claude — add provider capability matrix and supersede v1 surface doc

  • Added a provider capability matrix to root README.md (streaming / sessionResume / structuredOutput / ttft / gen / tail / tool) sourced from packages/polycli-runtime/src/registry.js. Surfaces minimax's session-resume / structured-output / streaming-timing limits in the root doc instead of only in plugins/polycli-opencode/README.md, and explains why qwen's tool: "missing" is semantically distinct from the other seven providers' tool: "unsupported".
  • Prepended a Superseded banner to docs/polycli-v1-public-surface.md (v0.3 snapshot) pointing to the live surface sources (README.md, registry.js, timing.js, CHANGELOG.md, docs/roadmap.md). Content below the banner is preserved as historical reference.
  • Updated CLAUDE.md Must-read list: added docs/roadmap.md, relabeled the v1 surface doc as a v0.3 snapshot / not a live constraint. Closes roadmap R3 (Path B) and R2 (minimax matrix).

2026-04-24 — Claude — v0.4.1 shipped

  • Pushed main + tag v0.4.1 (HEAD at eaddbbd).
  • GitHub release live at https://github.com/bbingz/polycli/releases/tag/v0.4.1 (notes from docs/release-notes-v0.4.1.md).
  • npm: @bbingz/polycli-opencode@0.4.1 published. Earlier suspicion that v0.4.0 had never reached npm was a token-expired false positive; npm view confirms 0.3.0 / 0.4.0 / 0.4.1 all present.
  • Closes the review loop that began with docs/archive/review-2026-04-24.md. All 8 bugs from docs/archive/review-2026-04-24-bugs.md (B1–B8) plus FU1–FU3 release-prep follow-ups landed. 221/221 tests pass at release tag.

2026-04-24 — Codex — lift ask response model to top level for all providers

  • Six provider runtimes (claude / gemini / kimi / minimax / opencode / pi) now populate the top-level model field on ask results, matching qwen and copilot.
  • Added a defaultModel pass-through in the registry so the host companion can supply a cached getAuthStatus value as a final fallback when the provider's own events do not carry a model.
  • Per-provider fixture and fallback assertions lock the new contract; integration smoke tests assert non-null model for all 8 providers.

2026-04-24 — Codex — fix observed host-companion bugs

  • Normalized provider CLI availability details to the first non-empty line so multi-line version banners no longer break setup text rendering.
  • Made --json errors structured for argument and lookup failures, added subcommand --help short-circuiting, and validated timing --provider / --history inputs.
  • Aligned cancel no-op exit behavior and flattened result --json so completed job payloads expose response, ok, sessionId, and timing at the top level while retaining job metadata.

2026-04-24 — Codex — prepare v0.4.1 release

  • Bumped the four host plugin release manifests and host marketplace metadata from 0.4.0 to 0.4.1, while keeping internal workspace packages on the 1.0.0 line.
  • Marked @bbingz/polycli-utils and @bbingz/polycli-timing as private internal bundler inputs, matching @bbingz/polycli-runtime.
  • Removed the Copilot adapter from the Claude marketplace now that Copilot has its own marketplace, and kept release manifest validation wired into release:check.

2026-04-24 — Codex — fix pi duplicate terminal text and pin default model

  • Fixed the pi JSON stream parser so visible assistant text is sourced from text_delta when present, with terminal message_end / turn_end / agent_end text only used as a fallback; this removes the repeated final-answer concatenation seen in live pi runs and saved fixtures.
  • Pinned the default pi model to openai-codex/gpt-5.4 so host-driven ask calls no longer depend on whatever ambient pi config happens to select.
  • Updated pi fixture expectations and added focused regression coverage for both the duplicate-terminal-summary case and the default-model invocation contract.

2026-04-24 — Codex — harden ask-path visibility for qwen and kimi

  • Added prompt-runtime constraints for prompt-bearing commands so ask now applies the missing host-level visibility guards: kimi gets --no-thinking --max-steps-per-turn 1; qwen gets --max-session-turns 1 plus a forced visible-final-answer system prompt.
  • Added focused regression coverage for those ask-path constraints and the shared helper that computes them, closing the gap where /review was hardened but /ask was still unconstrained.
  • Extended fake provider integration fixtures so qwen/kimi can simulate thought-only output when constraints are absent; this makes the failure mode observable in CI instead of only via live session history.
  • Verification on main: npm test passed 191/191; live non-Claude smoke (gemini, kimi, qwen, minimax, copilot, opencode, pi) all returned successful setup + ask, with qwen/kimi now producing visible final OK through polycli.

2026-04-22 — Claude — v0.4.0 shipped

  • Pushed main + tag v0.4.0 (HEAD at 08f84c3).
  • GitHub release live at https://github.com/bbingz/polycli/releases/tag/v0.4.0 (notes from docs/release-notes-v0.4.0.md).
  • npm: @bbingz/polycli-opencode@0.4.0 published; npm view confirms no propagation lag.
  • Closes the review loop that began with docs/archive/review-2026-04-22.md. All P0/P1 from that review plus P3 fixture migration landed. 185/185 tests pass at release tag.

2026-04-22 — Claude — Group 4 / 5 / release specs for Codex

  • Appended three spec sections to docs/archive/review-fb64b1e.md:
    • Group 4: P2 host-plugin hygiene (appendPreview O(n²), previewText emoji, auto-scope shallow-clone diagnostics) — one-commit batch.
    • Group 5: real-CLI saved-stdout fixture migration (per-provider capture list, replay helper design, scope guards against running capture in CI).
    • Release checklist: v0.4.0 step-by-step with explicit Codex / user / Claude role split for each step.
  • Created tasks/lessons.md with the "stay in reviewer role" correction — do not slip into implementation mode when the division of labor is Claude-reviews-Codex-implements.

2026-04-22 — Claude — Verdict on commit 95b003c (Group 2 + 3 complete)

  • Appended verdict A to docs/archive/review-fb64b1e.md: atomic-save durability + /review CLI hard constraints both landed; npm test 171/171.
  • Codex's Phase 1 research (docs/archive/review-cli-flags.md) corrected 4 of 6 CLI-flag hypotheses from the original review against locally installed CLIs + primary sources.
  • Non-overridable review constraint decision documented and enforced via assertNoReviewConstraintOverride.
  • Original review P0/P1 scope now 100% closed. P2/P3 move to release backlog.

2026-04-22 — Claude — Verdict on commit 6636b7a + Group 2/3 instructions

  • Appended verdict to docs/archive/review-fb64b1e.md: Group 1 landed cleanly (P1-C / P1-D / P1-E / registry gemini branch), 152/152 tests passing.
  • Group 2 (atomic-save durability) and Group 3 (/review CLI hard constraints) now have concrete per-file fix specs appended to the same doc, including test plans and scope guards.
  • Group 3 requires a Phase 1 research pass (per-provider CLI flag verification) before Phase 2 code changes; suggested output is a short memo at docs/archive/review-cli-flags.md.

2026-04-22 — Claude — Follow-up review of commit fb64b1e

  • Authored docs/archive/review-fb64b1e.md: verdict A- on Codex's fix batch.
  • All 6 P0 items fixed with correct semantics + 23 new regression tests (npm test: 146/146).
  • P1 fixed: A (transient probe), B (session-id fall-through), G (jobs/.json locking), H (stdout-as-error paths).
  • P1 deferred for next batch: C (generic event.text fallback), D (copilot/opencode type:"error" capture), E (gemini hasVisibleText), F (atomic-save fsync), I (/review CLI hard constraints), plus registry isTerminalSummaryEvent gemini branch.
  • Recommended next-commit grouping: Group 1 (streaming parser consistency = C+D+E+registry); Group 2 (atomic-save durability = F); Group 3 (/review constraints = I).

2026-04-22 — Claude — Full implementation review (for Codex handoff)

  • Authored docs/archive/review-2026-04-22.md: 4-agent parallel review across utils+timing, runtime core, 8 provider adapters, and host plugin (~5,200 LoC source, ~117 raw findings).
  • Report is structured as P0 (6 ship blockers) → P1 (9 high-risk themes) → P2 (parser / timing / process / host grouped) → P3 (gaps and nits).
  • Flagged one runtime-core agent recommendation to REJECT: timing.js:66, 73 hardcoding cold / retry as unsupported is correct per the documented project decision; do not change to missing.
  • No source code touched; review doc only.

2026-04-22 — Claude — Repo onboarding scaffolding

  • Added CLAUDE.md at repo root: thin Claude-Code-specific patches layered on top of AGENTS.md (architecture boundary, test command priority, provider gotchas).
  • Added this CHANGELOG.md to satisfy cross-AI collaboration convention from the user's global rules.
  • Seeded project memory under ~/.claude/projects/-Users-bing--Code--polycli/memory/ with MEMORY.md index and layered entries (user / project / feedback / reference).
  • No source code touched; no tests run (docs/infra only).