mirror of
https://github.com/Dicklesworthstone/coding_agent_session_search.git
synced 2026-09-22 11:59:25 +08:00
master
3 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Dicklesworthstone
|
5f05938484 |
feat(pages,doctor,status,indexer,sources,gate): pages key CLI, WAL hygiene, segment pressure, setup sync, install fall-through, docs validator, and end-to-end proofs for #439/#440/#441 (coding_agent_session_search-ctigq, -2w1sc, -hr9rn, -q2vn9, -z2uon, -wpe6s, -vg2e1, -tcrxi, -c2xxq, -xwuon, -jwox6, -vdyxd, -apriu)
Correction to 1f642a24's message: the golden leak was not caused by the golden harness's own stale-on-read refresh — verify7 reproduced it with that opt-out in place. The writer is tests/cli_robot.rs: 52 call sites use tests/fixtures/search_demo_data IN PLACE through a base_cmd() that never disabled auto-refresh, and on a fleet worker the checkout is not under the OS temp dir, so a robot-CLI search spawned `cass index --background` into the committed fixture and ingested that worker's real sessions before the golden tests copied it. base_cmd() now sets CASS_AUTO_REFRESH=0. Proof: verify9 ran the in-place robot tests before golden regeneration in one job and search_robot/stats_full_payload came back byte-identical to HEAD. pages key (ctigq): `cass pages key list|add-password|add-recovery|revoke| rotate --archive <bundle>` over the complete key_management engine (3,915 lines, zero callers until now). Passwords via prompt or --password-stdin (current on line 1, new on line 2), never argv; --json; exit 1 engine refused, 2 usage, 3 not a bundle, 6 password required. Four unit tests on a real encrypted bundle plus a CLI suite (tests/pages_key_cli.rs). docs/RECOVERY.md rewritten to the real surface; README row. WAL hygiene (z2uon): forget --apply, dedup --apply and analytics validate --fix close through the index run's checkpointing close; doctor `archive_wal` (pass ≤ 64 MiB; warn with size and remedy; warn-and-defer under an index run; --fix checkpoints and reports pass / blocked / failed truthfully) with a sparse-WAL e2e in tests/cli_doctor.rs. status/health e2e for db_bytes / wal_bytes / shm_present, now on the `status` surface too (q2vn9). Segment pressure (wpe6s, GH #441): quill_bridge::segment_file_count (metadata-only upper bound), status/health `index.segment_files`, doctor `index_segments` (warn above 8× the merge threshold, `cass index --full` remedy), CASS_TEST_SKIP_POST_RUN_LEXICAL_MAINTENANCE hook and an e2e that builds a 40-segment generation, sees the warning, and proves a plain `cass index` consolidates it. GH #439 (vg2e1) and GH #440 (tcrxi) end-to-end proofs: FTS-repair hooks (sleep per page after the heartbeat; park before the first page) — a slow repair survives a 20 s abort window, a parked one exits 70 with the `index-stalled` envelope; a hook parks a force-rebuild between the staged engine commit and the checkpoint write, records the gap, and a plain `cass index` resumes with exit 0 and no duplicate identities. sources (vdyxd, G.1): setup no longer marks the final sync complete without running it — interactive setup runs `cass sources sync` afterwards and records it only on success; --json setup defers it and reports sync.status=pending. Remote install walks candidate_methods() (binstall → verified prebuilt → cargo install → bootstrap) and falls through on failure instead of stopping at the first choice. `cass sources sync --all` exists (jwox6): doctor, robot-docs and the fleet rehearsal recommended it while clap rejected it. installer (apriu, G.2a): install.sh probes host glibc and falls back to build-from-source below 2.38; proven on ubuntu:22.04 (2.35 → source route) and ubuntu:24.04 (2.39 → prebuilt). docs validator (xwuon, A.9): scripts/validate_docs.sh --keys/--flags/--env/ --truth check README key bindings against the TUI key map (123/123, planted fake key fails), every `cass … --flag` usage against `cass … --help`, and every README env var against the code; found and fixed `sources sync --all`, `timeline --days`, the `<mark>` highlight claim and a dead CASS_UI_METRICS row. gate.sh gains an opt-in docs-truth stage. gate (c2xxq): scripts/gate.sh had a false green — verify8 was cut by the fleet's 30-minute SSH ceiling after fmt+clippy and the script still said GREEN. Now a non-zero rch exit is RED, every expected stage (plus a terminal job-complete marker) must report, the receipt captures both streams and survives a RED run, --lib-only/--no-goldens split runs under the ceiling, the lib stage runs under `timeout`, and the fleet target dir no longer inherits the shell's CARGO_TARGET_DIR. contract (hr9rn): introspect/capabilities has_json_output looks through the subcommand tree (11 commands flip to true and gain the robot-format arg). TUI: the unreachable `Alt+W => swarm cockpit` arm removed. CHANGELOG [Unreleased], AGENTS swarm/issue rules, plan §7b progress log. Verification (fleet, scripts/gate.sh STAGE receipts, RUST_MIN_STACK=16777216): - verify9 (2026-09-02 00:35-00:58): fmt 0, clippy 0, goldens-regen 0, goldens 0 with search_robot/stats_full_payload byte-identical to HEAD after the in-place cli_robot tests ran first (leak fix proven); cli_index/cli_robot/cli_status 101 (Codex fixture names, thresholds, health nesting) fixed in tree and re-run below. - verify10 (01:04-01:19): `cargo test --lib` to completion, 6,847 passed / 6 failed / 38 ignored, no wedge; the six (bookmarks missing from CANONICAL_TOP_LEVEL_COMMANDS, four sqlite3-CLI tests on a worker without sqlite3, daemon socket path over sun_path on a long TMPDIR) fixed in tree and re-run below. - verify11 (05:08-05:23): fmt 0, clippy 0, lib-tests 0 (canonical commands, daemon, sqlite, sources install/setup, pages key, quill_bridge, stall diagnostics, fingerprint cache), pages_key_cli 0, cli_robot 0, cli_doctor 0, tui_flows 0, bookmarks_cli 0, goldens-regen 0, goldens 0; cli_index 101 (gh439 reads the last stdout JSON line now) and cli_status 101 (health nests state.database) fixed in tree. - verify12 (05:41-05:47, full lib suite on this tree): fmt 0, clippy 0, lib-tests 0 — 6,855 passed / 0 failed / 38 ignored in 144 s. - verify13 (05:47-05:55, --regen-goldens): fmt 0, clippy 0, lib-tests 0 (quill_bridge, stall diagnostics), cli_status 0, cli_doctor 0, pages_key_cli 0, goldens-regen 0, goldens 0; cli_index 101 because the gh439 slow-repair test parsed the last LINE of a pretty-printed summary (harness bug) - it now stream-parses the last JSON document. - verify14 (05:57-06:00, cli_index gh439 after that fix): fmt 0, clippy 0, test-cli_index 0 (gate: GREEN). Goldens: 12 reviewed diffs (segment_files / db_bytes / wal_bytes / shm_present fields, has_json_output flips, pages key and sources sync --all in the command trees, doctor archive_wal / index_segments checks); the four that differ only by the regeneration writer's dropped final newline were left at HEAD. Measurement (owner's archive: 10.3 GB db, 200 MB WAL, 267 Quill segment files; `search "borrow checker lifetimes" --robot --robot-meta --limit 5`, five warm runs, CASS_AUTO_REFRESH=0): v0.7.1 median wall 15.3 s (search_ms 6.3-7.5 s, other_ms 1.0-1.9 s); this tree (release build) median wall 1.9 s (search_ms 337-406 ms, other_ms 754-810 ms after a first-run 2.6 s sidecar miss). hits=1 on both. Not yet isolated: the engine-time drop happened with the segment count unchanged, so the per-open seal in the old Quill config is the likelier cause than fragmentation on this archive. Consolidation of the owner's 267 segments was NOT achieved: the documented path (`cass index`, and `cass doctor --fix` for the WAL) never leaves the frankensqlite writable open on that archive - one core, 6.7 GB RSS, ~30k 24-byte header reads/s on the 200 MB WAL, no progress in 7 min - while read-only opens take 1.5 s. On a COPY whose WAL C SQLite checkpointed losslessly (1,012 conversations / 538,807 messages before and after), the same `cass index` gets through the open and indexes, so the WAL is the trigger; C SQLite's quick_check on that copy also reports a doubly- referenced freelist page and free-space corruption on ~35 pages that `cass doctor` calls healthy. The archive's last completed index is 2026-08-14. Recorded as a new P0 in the plan (section 7b); the live archive was not modified. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> |
||
|
Dicklesworthstone
|
c59c5da9b8 | fix(scripts): rch-wrap docs validation cargo | ||
|
Dicklesworthstone
|
3ecbf19ca9 |
Add documentation testing infrastructure (P6.12)
- tests/docs/readme.rs: Tests for README.md generation accuracy - tests/docs/help.rs: CLI help and generated HTML validation tests - scripts/validate_docs.sh: Shell script for link and content validation - docs/DOCUMENTATION_STYLE.md: Documentation style guide - Cargo.toml: Added [[test]] entry for docs test crate Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> |