Release Notes¶

Highlights since the v6.0 Knowledge Context Injection milestone (commit 0fdb7665). Grouped by subsystem.

Knowledge Injection — Copilot per-turn injection is now version-adaptive multi-channel (2026-07-19)¶

GitHub Copilot previously received only a session-start baseline; per-turn injection was blocked because Copilot filesystem hooks can only inject via an additionalContext field, and which event's additionalContext is honored changes between Copilot versions (postToolUse on ≤ 1.0.71; userPromptSubmitted newly on 1.0.72+, where postToolUse went flaky). A fixed channel was never upgrade-safe.

src/hooks/copilot-channel-capabilities.js now maps the installed Copilot version to the set of channels to emit on, and that set is the deduplication:

Retrieval runs at most once per turn (block cached in the session stash; the fail-safe second channel reuses it). Overrides: COPILOT_KB_CHANNELS (force set / none), COPILOT_VERSION (test). scripts/verify-copilot-hook-injection.sh reports which channel delivers after an upgrade, to keep the map honest. Resolver + end-to-end tests under tests/experiments/copilot-*.test.mjs.

See Knowledge Context Injection.

Observational Memory — SQLite → km-core cutover complete (2026-06-05)¶

Phase 44 Plans 12 → 18 finished cutting every production code path off the legacy .observations/observations.db SQLite file. The file was archived to .observations/observations.db.archived.2026-06-05 after Plan 44-18 Tasks 1-4 landed (commits cf6c8da45 → c837dc421).

The deferred Plan 44-12 § "fully unused observations.db" success criterion — carried through 44-13, 44-14, 44-17 — is now honored. obs-api keeps running cleanly with no SQLite handle; the DB_PATH constant was dropped from the script in the Task 5 cleanup commit. The archived file is preserved in .observations/ (gitignored under *.db.archived*) for one-shot historical inspection should a regression need pre-cutover data.

Constraint System¶

A run-through of the persisted violation-history.json showed that 9,387 of 9,780 lifetime violations (96 %) were no-magic-numbers — its \b\d{2,}\b pattern matched any 2+ digit number including port numbers, PIDs, and Bash command digits. Critical-severity rules like no-hardcoded-secrets had fired exactly once in the same window. A second sweep uncovered a config split-brain: host hooks loaded ${CODING_REPO}/.constraint-monitor.yaml while the in-container dashboard loaded a stale integrations/mcp-constraint-monitor/constraints.yaml. Same rules diverged across the two — no-console-log ran as error on the host but warning in the dashboard.

See Constraint System.

Health Monitoring¶

Phase 33 follow-on: statusline + ETM pipeline (May 8)¶

Five intertwined bugs surfaced during a single debugging pass after the Phase 33 coordinator cutover. All fixed; published-doc set rewritten.

See Health Monitoring, Status Line, and the updated health-monitoring-overview.png / health-coordinator-architecture.png diagrams.

host.docker.internal rewriting¶

The host-side health-verifier daemon kept reporting the LLM CLI proxy "unavailable" because host.docker.internal doesn't resolve outside Docker (the rule is shared with the in-container verifier, where it works). checkHTTPHealth now rewrites host.docker.internal:* → localhost:* when running on the host (/.dockerenv not present).

Bind-mount staleness supervision¶

macOS Docker Desktop occasionally caches single-file bind-mounts and stops reflecting later host edits — what the dashboard saw inside the container was a truncated snapshot of server.js (195901 bytes vs 200086 on the host), which broke startup with a syntax error mid-line.

A new verifyBindMountFreshness check compares host stat vs docker exec stat for each watched file. On size mismatch it raises a bind_mount_freshness violation and the refresh_bind_mounts remediation runs docker-compose up -d --force-recreate coding-services.

Watched files: .constraint-monitor.yaml, .global-lsl-registry.json, server.js, consolidate-observations.js. Add more in config/health-verification-rules.json under services.bind_mount_freshness.files.

See Health Monitoring.

Observational Memory¶

Single-owner observations gateway (May 3)¶

Periodic observations.db.corrupted-* files (10+ in 8 days) were traced to the canonical SQLite-on-Docker-Desktop-Mac corruption pattern: the host transcript monitor and the in-container dashboard both opened the file across the bind-mount boundary, and Apple Virtualization VM bind-mounts don't provide mmap coherence for SQLite's WAL/SHM shared memory.

A new host service — the Observations API server at scripts/observations-api-server.mjs (port 12436) — is now the single owner of observations.db. Every other consumer reaches the DB exclusively through this HTTP gateway:

Pattern mirrors the OKB/VKB approach: one process owns the DB, everyone else over HTTP. Same five-phase rollout structure (build, migrate reads, migrate writes, migrate consolidation, drop bind mount). Verified by zero new corruption files in the 48 h window after cutover.

Per-project consolidation (Apr 26)¶

Phases A/B/C/D made the pipeline project-aware end-to-end:

• Observations carry a project column populated by the LSL classifier
• ObservationConsolidator partitions observations by project before LLM grouping
• Insight synthesis runs per project — no more cross-project narrative bleed
• Digests/Insights pages have a project selector; API accepts ?project=<name>

Consolidator heartbeat + orphan sweep (Apr, superseded May 3)¶

Detached spawn originally protected the SQLite WAL across dashboard restarts but left orphans immortal — a 21-hour 0%-CPU run on the live dashboard prompted the work. The CLI wrote .observations/consolidation-heartbeat.json every ~2 s; any stderr line refreshed it; the dashboard swept stale heartbeats on startup.

This whole arrangement was superseded by the single-owner gateway above: consolidation now runs in-process inside the host obs-api server, which already owns the SQLite handle. No spawn, no second writer, no orphan-child timeout enforcement to worry about. The heartbeat file is still written (now by the obs-api itself) and surfaced via /api/consolidation/status so the dashboard's progress indicator continues to work.

Mixed-topic safeguards in the knowledge graph¶

A bug walk-through discovered a "GSD Statusline and Hook Integration" entity that bundled a GSD changelog with an LSL tmux-indicator description. Cause: fuzzy name dedup at Jaccard ≥ 0.7 with no content check — hook + integration overlap was enough to trigger a merge.

Three guards now run in persistence-agent.ts:

Sanitizer + redaction-pattern correction¶

• ObservationSanitizer repairs legacy <AWS_SECRET_REDACTED>frag corruption using sibling fields as the recovery oracle (e.g. matches by basename across modifiedFiles arrays).
• The aws_secret_standalone regex got lookarounds so a 40-char run inside a base64-like path can no longer be eaten as a "secret".

See Observational Memory.

Token Usage¶

Phase 36: per-(window, user) hourly exports + model-name canonicalization + treemap hover (May 16)¶

The Token Usage dashboard had two structural problems and one cosmetic one. All three closed in Phase 36 (7 plans).

Per-(window, user) hourly exports. The single .data/llm-proxy-export/token-usage.json blob ballooned past 600 KB / 1457 rows after ~24 h on a single contributor and would have caused git merge conflicts as soon as a second user pushed exports. Switched to the same filesystem convention LSL uses: YYYY/MM/YYYY-MM-DD_HHMM-HHMM_<hash6>.json. The time-window string is now sourced from the health coordinator's new /health/state.lsl_meta.current_window field (single source of truth, computed by getTimeWindow(utcToLocalTime(now))); the proxy fetches with a 30 s cache + local fallback. The 6-char user-hash is exported by the wrapper script (_work/rapid-llm-proxy/bin/start-llm-proxy.sh) as LLM_PROXY_USER_HASH before exec node, derived from scripts/user-hash-generator.js — same hash logic used everywhere else in the project so it cross-machine-reproduces deterministically.

Cross-user merge contract. A user_hash column and UNIQUE INDEX (user_hash, id) discriminate rows. hydrateFromExports() runs on every proxy boot (the old count > 0 → return early-exit is gone) and recursively walks <baseDir>/**/*.json, inserting with ON CONFLICT(user_hash, id) DO NOTHING. After git pull brings down a peer's ..._<other-hash>.json, the next proxy kickstart ingests it and the peer's rows appear in the dashboard alongside yours. The composite key supplies idempotency, not skipping.

.gitignore cleanup. .db-wal / .db-shm / .db-journal weren't covered by the old *.db rule, so .data/llm-proxy/ kept showing dirty in git status whenever the SQLite WAL had data. Replaced the single *.db line with explicit per-suffix entries matching the existing .data/knowledge.db precedent. !.data/llm-proxy-export/ allow-list preserves the new per-hour files for git.

Model-name canonicalization. The "By Model" panel was showing 8 rows for what is really 3 Claude models — each provider returns its own spelling and the proxy was recording the raw string verbatim. canonicalizeModelName(raw) now lives next to the existing model-maps in proxy-bridge/server.mjs and is applied once at the logTokenCall site. The raw spelling is preserved per row in a new model_raw column (same PRAGMA-guarded ALTER pattern as user_hash) so debugging "did Copilot serve the dated snapshot or the rolling alias?" still works. An idempotent backfill on proxy init rewrites pre-existing rows once (WHERE model_raw IS NULL); re-runs are no-ops.

Treemap hover tooltip. The "Hover for details" subtitle on the Token Consumption by Process treemap was aspirational — recharts.Treemap had no <Tooltip> child, and the inline label only renders for boxes ≥ 40×30 px so smaller processes were silent. Added a TreemapTooltip component (process / total / in/out split / calls / avg latency) wired as a Tooltip child, plus an SVG <title> element inside each rect for screen-reader/native-browser fallback. Playwright-verified on both small (reap-final, 81×44 px) and large (observation-writer, 981×352 px) boxes.

See Token Usage and the updated token-usage-architecture.png / health-mon-tokens-usage.png diagrams.

Dashboard / VKB rendering¶

Other¶

• coding-services Docker config: CODING_REPO=/coding env var; bind-mounts for .constraint-monitor.yaml, .mcp-sync/, the dashboard server.js, and the consolidator CLI.
• gsd-update workflow: bumped to GSD 1.38.5 after the previous run; 7 custom skills (codecraft-zuul-api, ddad-rpu, find-docs, fix-vulnerabilities, get-codeowner, raas-api, s3-session-search) were preserved across the install.
• Cross-team relations + project-anchor connection in the KG so insights from online learning attach to the right project node.