System state
Source maintenance checkpoint — 2026-10-29
Inspected baseline: main@471ff6c,
the merge of #577. The canonical dispatcher
reports 1.8.1. This checkpoint measures source and reconciles issue state; it does not
remeasure an installed home, native consumer, trust decision or live canary.
ACTIVE owns the current work queue. ROADMAP owns remaining dependencies. All earlier status text, including dated runtime evidence, is preserved in the October archive.
Source already on the baseline
| Surface | Accepted source and evidence | Limit |
|---|---|---|
| Learning and replay records | #544 validates complete outcome streams; #562 refuses nonfinite/non-JSON publication; #565 adds recursive string-key validation and strict advisory-ledger JSONL. #558/#563 are closed. | Trusted ingestion, omitted-tail detection and event identity remain #511. The advisory ledger is distinct from the learning stream; its line size remains uncapped. |
| Sync identity and cost | #559 compares selected-tree identities in both directions; #566 reuses snapshots within one preflight and retains fresh apply revalidation. #556 is closed. | Snapshots do not exclude concurrent writers. The later native bind-mount proof is recorded in #587 below. |
| Replay publication | #555 privately stages output and binds caller parents; #560 checks named inputs, visited transcripts and fellow outputs. | Existing target security and scanner limits remain; no multi-output crash-atomicity claim. |
| Canary chronology | #548 excludes observations before the candidate/genome cutoff. | Labels and producer timestamps do not prove enrolled execution. #520/#517 retain the trusted backend boundary. |
| Replay regressions | #577 pins duplicate-decision and tree-symlink refusal behavior. | The later macOS acknowledgement failure is tracked in #578 and addressed by candidate #581. |
Verified merge receipts: #565 is 21b35d1012d4aa7ffca754bf58a6f8f6a0bea560; #566 is
c2da76cf4912d4e379e93bfd6cfce8fb65abd890. Both merged on 2026-10-14. They supersede the
2026-10-13 pages’ open claims for #556/#558/#563; their historical records remain intact.
Reviewed maintenance work
The source PRs below have independent reviews with no confirmed findings. Windows CI on #580 exposed collision-message and CRLF-fixture test assumptions; corrections preserve refusal and exact source-byte checks. Use each PR’s current head, checks, review receipts and merge result when integrating; another branch’s green suite does not prove that tree.
| PR | Behavior under review | Remaining acceptance |
|---|---|---|
| #579 | System comparisons become explicitly observational and emit no canary gate. Promotion consumers reject old system-run authority. Regression weighting and triage deduplication are corrected; compatibility bundle advances to v5. | #520/#517 need trusted assignment/execution/store evidence. #511 needs trusted ingestion offsets and event-ID/cursor migration. Follow the PR’s archival-read and derived-data rebuild contract. |
| #580 | Bundle failures name recovery bytes and component/target mappings. Native name-only lookup checks absent destination and recovery spellings after identity/device preflight. | Carries #278’s closure. The local managed environment cannot run mount qualification; #587 supplies the native CI proof. |
| #581 | Atomic fixture acknowledgement, honest POSIX group cleanup, isolated Windows supervisor startup, bounded regex replies/stderr and execution-contract identity. | Carries #578/#576’s closure. Native CI supplies Windows/macOS/Linux evidence. |
| #582 | Tests own their fixture roots and match the Codex executable basename. | Local root suite: 2,114 tests, 68 skips; affected modules: 501 tests, seven skips. These results belong to #582’s head. Native CI remains the platform gate. |
| #583 | Recorded 50-case replay lab, visual docs entry point and bounded public asset staging. The browser displays recorded comparisons. | 102 Python staging/link tests, four JavaScript tests, Pages deployment and desktop/mobile browser verification passed. Filtering never changes the 50-case gate denominator; no live execution or canary claim. |
| #584 | Platform/quote-aware shell anchors and reuse of one canonical-reference result, including UNPROVEN. | Carries #78’s closure, including wrapper-variable and duplicate-probe follow-ups. Independent review ran 59 tests: 58 passed, one native-Windows skip; native CI supplies platform proof. |
| #587 | Real same-device bind aliases refuse before probes or writes; a disjoint bind-mounted home installs successfully. | Native job 123247166458 passed seven tests with zero skips at head 63bc177. Carries #523’s closure. Full CI and merge receipts own source acceptance; no concurrent-writer exclusion is claimed. |
The combined maintenance tree before #587’s test-only addition ran 2,155 root unit tests with 70 expected skips and no failures. Local full smoke remains limited by read-only home fixtures; the native per-platform CI jobs supply full smoke evidence. The lab’s browser receipt records the deployed desktop, mobile, keyboard, filtering and gate checks.
Operating contracts: learning outcome validation, replay output boundary, replay input errors, test fixture startup, and SPECS.
Runtime and human evidence
Historical observation, 2026-10-13: Doctor reported source 1.8.1 versus installed 1.8.0 and an unproven worktree canonical comparison. That observation is preserved in the archive; it is not a measurement of this session’s runtime.
HUMAN_TODO retains H-18 (private Actions billing), H-19 (PyPI trusted publisher) and H-22 (adapter prerequisites and per-root trust/canaries, including extract-api and the other workstation). Its recorded deployment and canary history is dated evidence. A source merge, local invocation or static Doctor result cannot supply fresh-session runtime proof.
The dispatcher remains feature-frozen under BLUEPRINT. Source maintenance here does not extend its scope or complete the separate consumer rollout.
Authority, products and canonical homes
The 2026-09-29 owner decision permits merges after exact-head checks, global law 2 review and the three-minute aging floor, using merge commits only. No squash. Re-read the declared tier before mutation; SPECS and the global laws own the gate.
Owner-authored repository content is GPL-3.0-only, with the curl-derived fixture under its
own licence (LICENSING, third-party notices).
The approved clean charter-replay extraction is Apache-2.0; its manifest is path-pinned by
the 2026-10-04 decision. REPLAY_TOOL_PRODUCT owns that scope;
blueprint-plugin extraction remains deferred. Record any
change of source direction before the harness and extracted product diverge.
Measurements live in BENCHMARKS; frozen-floor limits live in FLOOR_LIMITATIONS. The workspace receipt lab and handoff profile are bounded reference contracts. Their synthetic evidence does not establish native adoption or authenticate external receipts.
Evidence terms: implemented means accepted source exists; deployed means destination bytes were measured; runtime-proven requires the applicable live observations; experimental limits the compatibility claim; historical evidence applies only to its recorded date and scope.