olp/docs/adr/0008-dashboard-and-audit-query.md

ADR 0008 — Dashboard + Audit Query Layer (Phase 3)

• Date: 2026-05-25
• Status: Accepted (D48, design-only — implementation D-days D49–D54 follow; Phase 3 close = v0.3.0)
• Authors: project maintainer (with AI drafting assistance)
• Related:
  • OLP v0.1 spec § 4.6 (Dashboard requirements — port from OCP with multi-provider support) and § 4.7 (observability endpoints)
  • ADR 0007 § 12 (Phase 3+ out-of-scope: Dashboard, audit query layer, rotation) — this ADR opens those deferrals
  • ADR 0007 § 13 (Option 3 hybrid migration to SQLite) — explicitly NOT triggered by Phase 3; in-memory ndjson scan is the v0.3.0 query model
  • ADR 0007 § 7 (Identity classes) — Dashboard auth gating reuses owner-vs-non-owner pattern (Dashboard is owner-only)
  • ADR 0007 § 8 (Audit ndjson schema) — the data source the query layer reads
  • ADR 0004 Amendment 2 (soft triggers deferred to v1.x) — quota panel sources from provider.quotaStatus() which is a contract method; per-provider returns what it can or null
  • D45 reviewer P2 deferral on tried_providers semantics — addressed in this Phase as D53 (separate D-day; not in ADR 0008 scope)
• Phase 3 kickoff authority: maintainer "go" + standing-autopilot grant (~/.cc-rules/memory/auto/standing_autopilot_phase_2.md in cc-rules bf0ed9a — the grant explicitly excludes Phase 3+ as needing new authorization; the "go" supplied that)
• Lanes pinned by maintainer 2026-05-25 (Phase 3 kickoff brief):
  • Lane 1 Dashboard tech stack: A — static HTML + vanilla JS + fetch (no build step; matches OLP "no bundler" ethos)
  • Lane 2 Audit query model: A — in-memory scan of audit ndjson per request (O(N) per query; family-scale acceptable; SQLite deferred to Option 3 trigger per ADR 0007 § 13)
  • Lane 3 Audit rotation: B — daily rotation, files named audit-YYYY-MM-DD.ndjson (UTC date)
  • Lane 4 Refresh: A — page poll every 30s (no SSE infra introduction)
  • Lane 5 Dashboard scope: B — full per spec § 4.6 (quota + per-provider counts/cache/fallback last 24h + multi-provider spend trend last 30d + top fallback chains by trigger count)

---

1. Context

OLP v0.2.0 ships per-key audit ndjson at ~/.olp/logs/audit.ndjson (ADR 0007 § 8) but provides no aggregate query surface or visualization. The audit file grows unbounded, and operators must tail/grep to observe basic facts ("which provider served the most requests today", "what's my cache hit rate", "how often did the chain fall back to OpenAI"). Phase 3 closes this gap with:

1. lib/audit-query.mjs — an in-memory aggregator that scans ~/.olp/logs/audit-*.ndjson files in a configurable time window and returns shaped summaries.
2. server.mjs /v0/management/* endpoints — three owner-only JSON endpoints exposing the aggregate data: /v0/management/dashboard-data, /v0/management/quota, /cache/stats.
3. dashboard.html — a single static HTML file served from /dashboard (owner-only) that fetches the JSON endpoints, renders 4 panels, and polls every 30s.
4. Daily audit rotation — at UTC midnight (or on first append after a UTC-date change), the live audit.ndjson is renamed to audit-YYYY-MM-DD.ndjson and a fresh audit.ndjson opens. Cross-file queries handle the rolling 30-day window.

Phase 3 deliberately does not add a build step, a database, or per-key UI write surface. The first three are out of scope (Option 3 hybrid is § 13's forward path; SQLite has a Node-baseline blocker per § 11). The last is a security surface that warrants a separate review pass (Phase 4+).

Phase 3 is the natural home for OCP's dashboard.html port (v0.1 spec § 4.6 — "Port OCP's dashboard.html with multi-provider support"). OCP's dashboard was single-provider; OLP's is multi-provider, which is the substantive change. The HTML structure is otherwise lifted.

---

2. Decision

The five lanes above are normative. The deviation lattice they sit in:

Lane	Pinned	Rejected (why)
1. Tech stack	Static HTML + vanilla JS + fetch	SSR (Node template literals) — adds server-side render path; CSP harder. SPA — adds build step, violates ethos.
2. Query model	In-memory scan of audit-*.ndjson per request	SQLite indexed mirror — touches ADR 0007 § 13 migration → engines bump prerequisite. In-memory rolling aggregate — complex (per-write incremental + window expiry), correctness risk.
3. Rotation	Daily UTC rotation (audit-YYYY-MM-DD.ndjson)	No rotation — single file grows unbounded at family scale ~10–100 lines/day but a year is ~10–50k lines, still ok but no natural query unit. Size-based (100MB / keep 5) — equivalent complexity, query range less intuitive.
4. Refresh	Page poll every 30s	SSE push — adds server-side subscriber state; not justified at family scale. Static-at-load — UX too poor.
5. Dashboard scope	Full per spec § 4.6	Minimal (quota + cache + fallback only) — spec is already drafted; full is ~1 D-day more for trend + top-chains panels. Plus key-mgmt UI — security surface delta, Phase 4+.

Phase 3 does not introduce: a database, a build step, an SPA framework, SSE for dashboard, or web-side key management. Each is a future-phase concern (Option 3 hybrid; SSE for dashboard if poll latency becomes pain; key-mgmt UI after a security review pass).

---

3. Storage layout (~/.olp/logs/)

~/.olp/logs/
  audit.ndjson                    — live append target (Phase 2 D45)
  audit-2026-05-24.ndjson         — yesterday (after rotation)
  audit-2026-05-23.ndjson
  audit-2026-05-22.ndjson
  ...

Files are append-only (no in-place edits). Rotation atomically renames the live file and opens a fresh one. All files have mode 0600; the logs/ directory is 0700.

Retention policy at v0.3.0: unbounded by default (operator manages disk). A Phase 3+ amendment may add automatic retention (olp_audit_max_days config) once an observed operational need exists.

---

4. Audit query layer (lib/audit-query.mjs)

A new module that reads the rotated + live audit files in a date range and returns aggregate summaries. Per-call O(N) where N = total lines in the date range (family scale: thousands per day = trivial).

4.1 Public API

// Read all audit lines in [startMs, endMs); returns iterator of parsed events.
// Skips malformed lines (logs warn) so a corrupted day doesn't kill the query.
export function* readAuditWindow({ startMs, endMs, olpHome }): Iterator<AuditEvent>;

// Aggregate request shape over a window. Returns:
//   {
//     window: { startMs, endMs },
//     request_count, status_2xx, status_4xx, status_5xx,
//     by_provider: { [providerKey]: { count, cache_hit, cache_miss, cache_bypass, fallback_count } },
//     by_owner_tier: { owner: N, guest: N, anonymous: N },
//     by_path: { '/v1/chat/completions': N, '/v1/models': N },
//     median_latency_ms, p95_latency_ms,
//   }
export function aggregateRequests({ windowMs, olpHome }): RequestAggregate;

// Top-N fallback chains by trigger count in window. Returns sorted array:
//   [{ chain: ['anthropic', 'openai'], count: 42, first_seen, last_seen }, ...]
export function topFallbackChains({ windowMs, limit, olpHome }): FallbackChainSummary[];

// Daily series of request_count + latency_median over N days. Returns sorted array:
//   [{ date: '2026-05-22', request_count, median_latency_ms, by_provider }, ...]
export function spendTrendDaily({ days, olpHome }): DailySpendEntry[];

// Cache hit rate snapshot (in-memory cacheStore stats + audit-derived numerator).
// Differs from /cache/stats: that returns the live in-memory CacheStore stats;
// this is the audit-side derived rate over the window.
export function cacheHitRateWindow({ windowMs, olpHome }): CacheHitRateSummary;

4.2 Window semantics

windowMs is a duration ending at "now"; [now - windowMs, now). The implementation walks files whose date overlaps that range: today's audit.ndjson always; audit-YYYY-MM-DD.ndjson for each prior date in range. A line is included only if its ts (ISO-8601) falls in the window.

For the 30-day spend trend, windowMs = 30 * 86400 * 1000. The implementation buckets per UTC day and returns one entry per day, including days with zero requests (sparse-fill).

4.3 PII discipline

Per ADR 0007 § 8, audit events contain no message content, no response content, no raw tokens. The query layer relays only the schema fields. It MUST NOT introduce derived fields that reveal content (e.g., "first 50 chars of prompt").

4.4 Error handling

• Missing file → empty iteration (not an error).
• Malformed JSON line → log warn audit_query_skip_malformed + skip; continue.
• File-read error (EACCES, etc.) → throw to caller; the dashboard endpoint surfaces 500 with diagnostic message.

---

5. Audit rotation

5.1 Trigger

Rotation fires on the first append after a UTC date change. Implementation lives in lib/audit.mjs (extended at D49). On each appendAuditEvent call:

1. Compute today = new Date().toISOString().slice(0, 10) (e.g., '2026-05-25').
2. Read a module-scoped _currentDate cached at startup.
3. If today !== _currentDate AND audit.ndjson exists AND it is non-empty:
   • Rename audit.ndjson → audit-${_currentDate}.ndjson (the previous day's date).
   • Set _currentDate = today.
   • Continue with the append (new audit.ndjson opens via append-create).

The check is per-call (microsecond cost). The rename is the only filesystem heavy op and fires once per UTC day.

5.2 External-cron alternative (bin/olp-audit-rotate.mjs)

An auxiliary script is shipped at D52 for operators who prefer cron-driven rotation (e.g., to rotate exactly at 00:00:00 UTC rather than "first request after midnight"). The script does the same rename + state-bump logic but can be invoked from a host cron / launchd job. The in-server check remains as a safety net; both can coexist.

5.3 Concurrent-rotation safety

In-process: the rotation logic is wrapped in a per-process lock (Map<key='audit-rotate', Promise>) so two concurrent appendAuditEvent calls don't both attempt the rename. External cron + in-server check: the in-server check sees the rename has already happened (file with today's date already exists if cron beat it); the no-op fallback is "if audit.ndjson exists, append; else create + append" — POSIX semantics.

5.4 Renamed-file query path

The query layer (§ 4) walks audit-${date}.ndjson files for any date in the window. Today's file is always audit.ndjson (not renamed yet); yesterday + prior are date-suffixed.

---

6. Dashboard panels (per spec § 4.6, Lane 5 = B full)

dashboard.html is a single static file served from /dashboard. Renders 4 panels in a 2×2 grid:

6.1 Panel 1 — Per-provider quota / credit pool

For each loaded provider, calls provider.quotaStatus() (ADR 0002 Provider contract). Returns whatever the provider can report (e.g., Anthropic Plan limits remaining, Codex credit pool balance) OR null (provider opts out — Phase 2 mistral has no quota API).

Each row shows:

• Provider key + display name
• Quota remaining / quota total (or "n/a" if null)
• Last poll timestamp

6.2 Panel 2 — Per-provider request count + cache hit rate + fallback rate (last 24h)

Uses aggregateRequests({ windowMs: 86400 * 1000 }). One row per provider showing:

• Request count (total served by that provider, regardless of chain position)
• Cache hit rate (% of requests where cache_status === 'hit')
• Fallback rate (% of requests where fallback_hops > 0)
• 5xx error rate (% of requests where status_code >= 500)

6.3 Panel 3 — Multi-provider unified spend trend (last 30 days)

Uses spendTrendDaily({ days: 30 }). Shows a sparkline-style chart (vanilla SVG, no library) with:

• X axis: 30 daily buckets
• Y axis: request count per day (stacked by provider color)
• Hover tooltip: per-day breakdown by provider

Note: "spend trend" is the spec's term — at v0.3.0 we don't have provider-side cost integration (Anthropic Plan is flat-rate per Anthropic 2026-06-15 split per the learning memory). So "spend" is proxied by request count. A future ADR may add cost weights per provider when commercial cost-tracking lands.

6.4 Panel 4 — Top fallback chains by trigger count

Uses topFallbackChains({ windowMs: 86400 * 1000, limit: 10 }). Lists top 10 chains:

• Chain shape (e.g., anthropic → openai)
• Trigger count
• First / last seen timestamps

6.5 Refresh model (Lane 4 = A)

The dashboard sets a 30s setInterval that calls fetch('/v0/management/dashboard-data') + updates DOM in place (no full reload). Initial fetch on page load. The interval pauses when the page is hidden (via document.visibilityState listener) to avoid useless background polls.

6.6 Localhost-bound by default

The dashboard is served from the existing OLP HTTP port (default 4567 since v0.4.0 / D60; 3456 pre-v0.4.0) which is already bound to 127.0.0.1 per server.mjs startup (server.listen(PORT, '127.0.0.1', ...)). No additional binding logic. Remote operators access via SSH tunnel; ADR 0007 § 7 owner-only auth provides the per-request gate.

---

7. Server endpoints (D50)

All endpoints are owner-only per ADR 0007 § 7 (owner-tier validation through authenticate). Non-owner identities get 401 / 403 per existing patterns; anonymous (when allow_anonymous: true) gets 401 (these are management endpoints, not user-facing).

7.1 GET /dashboard

Serves dashboard.html. Owner-only gated. Content-Type text/html; charset=utf-8. Static file read once at server startup + cached in memory (small, no need to re-read per request).

7.2 GET /v0/management/dashboard-data

Returns the JSON payload the dashboard's 30s poll consumes. Shape:

{
  "generated_at": "<ISO-8601>",
  "window_24h": <RequestAggregate from §4.1>,
  "quota": [
    { "provider": "anthropic", "quota_remaining": 1234, "quota_total": 5000, "polled_at": "<ISO-8601>" },
    { "provider": "openai", "quota_remaining": null }
  ],
  "spend_trend_30d": <DailySpendEntry[] from §4.1>,
  "top_fallback_chains_24h": <FallbackChainSummary[] from §4.1>,
  "cache_stats": <stats from server.mjs cacheStore.stats() — global aggregate>
}

7.3 GET /v0/management/quota

Returns just the quota array (subset of dashboard-data; useful for scripted monitoring).

7.4 GET /cache/stats

Returns the live in-memory cacheStore.stats() shape. Planning authority is OLP v0.1 spec § 4.6 (which names /cache/stats explicitly); ADR 0005's Consequences/Mitigations paragraph (~ line 279) references it as the monitoring surface for per-(provider, model) cache hit-rate breakdown.

Shape gap to resolve at D50. The current cacheStore.stats() in lib/cache/store.mjs:320-350 returns { hits, misses, size, inflightCount } — global aggregate only, no per-(provider, model) breakdown. If D50 reveals the shape is insufficient for Panel 2 (per-provider 24h cache hit rate, currently sourced from aggregateRequests audit-side rather than cacheStore.stats), the dashboard endpoint is satisfied. If a future panel needs the per-(provider, model) breakdown that spec § 4.6 implies, D50 amends the store shape + an ADR 0005 amendment fires at that time. Phase 3 acceptance criteria do not require the breakdown.

7.5 Audit on management endpoints

All four endpoints append an audit row via the existing appendAuditEvent pattern. Path values are /dashboard / /v0/management/dashboard-data / /v0/management/quota / /cache/stats. The 30s poll generates 2880 dashboard rows per day per owner — manageable at family scale, but noted as a knob (a future amendment may suppress audit for these paths if they become noise-dominant).

---

8. Auth gating

Reuses ADR 0007 § 7 owner-vs-non-owner model + introduces a second gating mode.

Two gating modes (this ADR formalizes the distinction):

• owner_only_trim (Phase 2 / D46 model) — non-owner identities receive a 200 response with a trimmed payload (e.g., /health returns { ok, version } only). Used when the endpoint has a baseline payload that is safe to share with all identities and an enriched payload only for owners.
• owner_only_block (Phase 3 / D48 new) — non-owner identities receive 401 invalid_or_revoked_key (or 401 auth_required if no token). Used when the entire payload is sensitive and there is no safe baseline to share (Dashboard quota stats, fallback chains by trigger, etc. all reveal operational behaviour that should not leak to non-owner identities).

The four new endpoints (/dashboard, /v0/management/dashboard-data, /v0/management/quota, /cache/stats) are owner_only_block. /health remains owner_only_trim.

The owner_only_endpoints config gains four entries; the gating-mode distinction is implementation-side (the handler decides whether to trim or block based on the endpoint). Server startup defaults owner_only_endpoints to include /health + the four new ones (Phase 3 default; operator can opt-out per-endpoint via config). Pre-Phase-3 deployments with owner_only_endpoints: ['/health'] continue to work — the new endpoints will 401 for non-owner under that legacy config because the handler is owner_only_block-mode regardless of the config list (the config controls /health's trim/full toggle only; the management endpoints are not opt-out-able to a non-401 response).

401 shapes match Phase 2 / D45 pattern: JSON { error: { message, type } } with type: 'auth_required' or 'invalid_or_revoked_key'.

---

9. Failure modes + graceful degradation

Failure	Behavior
audit.ndjson absent (fresh install)	Empty arrays in all aggregates; dashboard shows "No requests in window" panels
audit-YYYY-MM-DD.ndjson corrupted (one bad line)	Skip line + log warn; continue; dashboard rendering unaffected
Provider quotaStatus() throws	Panel shows that row as quota: "error"; other providers' rows render normally
Dashboard /v0/management/dashboard-data query >5s	Dashboard JS shows "Loading…" with a timeout; second poll attempts after 30s
audit.ndjson rotation fails (rename EACCES)	appendAuditEvent warn audit_rotate_failed + continues appending to the un-rotated file; next call retries the rotation
File handle limit hit during 30-day query (many files open)	Query reads one file at a time (no parallel reads); never opens >2 simultaneously

The dashboard degrades visibly (per-panel error states) rather than failing whole-page.

---

10. Acceptance criteria

Implementation D-days (D49+) MUST land tests covering:

1. readAuditWindow correctly iterates events from today's audit.ndjson + N prior rotated files within the window.
2. readAuditWindow skips malformed lines without throwing; logs warn for each.
3. aggregateRequests correctly counts by provider + cache_status + owner_tier + path; correctly computes median + p95 latency.
4. topFallbackChains returns sorted by count descending; ties broken by first-seen timestamp ascending.
5. spendTrendDaily sparse-fills zero-request days; window respects UTC day boundaries.
6. Daily rotation — writing past UTC midnight renames audit.ndjson → audit-<yesterday>.ndjson and continues appending to a fresh audit.ndjson.
7. Cross-file query — a 30-day window with mixed rotated files returns correctly merged results.
8. Concurrent rotation safety — N concurrent appendAuditEvent calls during a UTC date change result in exactly one rename + all lines append to the correct file.
9. GET /dashboard returns 200 HTML to owner; 401 to non-owner. This includes the case where allow_anonymous: true AND no Authorization header is presented: the authenticate middleware produces an anonymous identity, the owner_only_block mode then rejects with 401 (per § 8 — anonymous is non-owner; management endpoints block, do not trim). When allow_anonymous: false + no header, 401 fires earlier at the authenticate middleware itself. Test must cover both cases.
10. GET /v0/management/dashboard-data returns 200 JSON to owner with all required fields populated.
11. GET /cache/stats returns 200 JSON to owner with the live in-memory cache stats shape.
12. Dashboard HTML smoke — fetched via test http client + parsed → has the 4 panel containers + 30s poll script; no JS console errors when loaded in a real browser (manual or playwright; manual is acceptable at Phase 3).
13. Audit on management endpoints — calling /v0/management/dashboard-data appends an audit row with path: '/v0/management/dashboard-data' and status_code: 200.
14. Graceful degradation — when a provider's quotaStatus() throws, the dashboard endpoint still returns 200 with that provider's quota row showing "quota_remaining": null and an error indicator.
15. PII guard — every aggregate query function asserts at the test level that returned data does NOT include any message content; the prompt/messages/response/content fields MUST NEVER appear in any output shape.

---

11. Forward path (Phase 4+)

Items deliberately deferred:

• SQLite migration (Option 3 hybrid) — trigger: query latency >2s on a typical owner session, OR Dashboard usage scales beyond family (>5 owners polling). Preconditions per ADR 0007 § 13: engines bump + CI matrix change as a separate prior PR.
• SSE push for dashboard live updates — trigger: 30s poll feels stale, OR operator wants real-time view of streaming requests. Reuses existing streaming infra from ADR 0005 Amendment 8 (v1.x streaming SF when it ships).
• Key-mgmt UI from dashboard — owner can create/revoke/edit keys from the web UI rather than CLI. Out of Phase 3 because (a) it adds a write surface to the dashboard requiring careful CSRF handling, (b) security review of the auth flow is non-trivial, (c) the CLI surface from D47 covers the same use cases.
• Cost weights per provider — once provider-side cost tracking is feasible, "spend trend" can show actual dollars. At v0.3.0 it's a request-count proxy.
• Audit retention / max-days policy — currently unbounded; operator manages disk. A Phase 3+ amendment adds audit_max_days config when an operational need emerges.
• Per-key dashboard views — owner sees aggregate; per-key drill-down is a future amendment.

---

12. Out of scope (explicitly NOT in Phase 3)

• Per-key per-provider auth artifact mapping (ADR 0007 § 12; Phase 4+).
• tried_providers schema semantics fix on key_no_provider_access 403 — D45 reviewer P2 deferral. Tracked as a Phase 3 implementation D-day (D53) but NOT part of ADR 0008; documented in ADR 0004 amendment or ADR 0007 § 8 amendment at D53.
• All ADR 0007 § 12 deferrals other than Dashboard + audit query layer + rotation.
• Externally-visible Dashboard (anything bound to 0.0.0.0 / public). Operator SSH-tunnels.

---

13. Phase 3 sprint shape

D-day	Deliverable	Type
D48	This ADR (0008 draft)	ADR-only
D49	lib/audit-query.mjs + Suite 23 unit tests	impl
D50	server.mjs /v0/management/* endpoints + /dashboard route + Suite 24 HTTP tests	impl
D51	dashboard.html + render JS + 30s poll	impl
D52	Audit daily rotation (lib/audit.mjs extension + bin/olp-audit-rotate.mjs + Suite 25 rotation tests)	impl
D53	tried_providers schema fix (D45 P2 deferral; small)	impl
D54	E2E browser smoke (manual or playwright) + AGENTS / README polish	tests + docs
D55	Phase 3 close → v0.3.0 (release-kit PR per phase_close_trigger)	release

Each D-day = implementor + fresh-context opus reviewer per Iron Rule 10. Estimated wall-clock: similar to Phase 2 cadence (1 intense session per D-day under standing autopilot).

---

Consequences

Positive:

• Operators get an at-a-glance view of OLP's behaviour (was a tail/grep exercise pre-Phase-3).
• Audit layer becomes queryable, not just append-only; lib/audit-query.mjs is reusable for future CLI tools (olp-audit search etc.).
• Daily rotation bounds per-file size + creates a natural archival unit.
• Owner-only gating reuses Phase 2 auth model (no new auth surface).
• v0.1 spec § 4.6 Dashboard requirements are met without introducing a build step / database / framework.

Negative / trade-offs:

• In-memory ndjson scan is O(N) per query; if audit grows to millions of lines (single-host years), the 30-day query becomes slow. Mitigation: ADR 0007 § 13 Option 3 hybrid is the documented next step; trigger is observed slowness.
• 30s poll generates baseline traffic when dashboard is open (2880 management requests/day per owner). Not a real cost but worth observing.
• Audit rotation is "first append after UTC midnight" which means a server with zero requests overnight rotates lazily (first request of the new day triggers it). External cron at D52 covers the strict-midnight case.
• No automatic audit retention. A multi-year-running server accumulates files; operator manages.

Reversibility:

• Dashboard is a static file + 4 endpoints. Removable in a future revert PR if Phase 3 retrospectively proves unwanted.
• Audit rotation is additive — disabling reverts to single-file behaviour without code change (operator never invokes the cron + the in-server rotation can be guarded by a config flag).
• The lib/audit-query.mjs module is consumed by Dashboard endpoints + can be used standalone; removing it requires unrelated endpoint surgery.

---

Authority citations

• OLP v0.1 spec § 4.6 + § 4.7 (Dashboard + observability endpoints) — at ~/.cc-rules/memory/projects/olp_v0_1_spec.md on the maintainer's workstations.
• OCP dashboard.html (prior-art reference for the multi-panel HTML structure) — at ~/ocp/dashboard.html on the maintainer's workstation; OCP production reference.
• ADR 0007 §§ 7 / 8 / 12 / 13 (owner-gating model; audit ndjson schema; Phase 3 scope opening; SQLite forward path).
• ADR 0002 (Provider contract — quotaStatus method that Panel 1 consumes).
• OLP v0.1 spec § 4.6 (planning authority for /cache/stats endpoint name + Dashboard requirements). ADR 0005's Consequences/Mitigations paragraph references the endpoint as the monitoring surface for per-(provider, model) cache hit-rate breakdown; that breakdown is a Phase 4+ amendment trigger if needed (see § 7.4 above).
• ADR 0004 Amendment 5 (D40 X-OLP-Fallback-Detail — top-fallback-chains panel data shape lineage).
• Standing-autopilot grant (~/.cc-rules/memory/auto/standing_autopilot_phase_2.md in cc-rules bf0ed9a) — Phase 3 kickoff via maintainer "go" + lane pin.
• Phase 2 kickoff handoff pattern (~/.cc-rules/memory/handoffs/2026-05-25-phase-2-kickoff.md in cc-rules d9da966) — this ADR follows the same structure.
• CLAUDE.md release_kit.phase_rolling_mode current_phase: Phase 3 — confirms this ADR lands in the Phase 3 sprint.