pi-smart-compact 8.0.8 → 9.1.0

package/ARCHITECTURE.md

@@ -60,14 +60,14 @@ threads a typed context through ten stages:
 
 | # | Stage | Module | Transition |
 | ---: | --- | --- | --- |
-| 1 | prepare | `app/steps/prepare.ts` | config + auth + provider caps |
-| 2 | window | `app/steps/window.ts` | pick the prefix to compact |
+| 1 | prepare | `app/steps/prepare.ts` | config + provider caps + budgets; auth remains lazy |
+| 2 | window | `app/steps/window.ts` | pick the prefix using a calibrated final-summary allowance |
 | 3 | recover | `app/steps/recover.ts` | restore log-truncated messages |
 | 4 | tier | `app/steps/tier.ts` | choose none / light / full |
 | 5 | extract | `app/steps/extract.ts` | prune + deterministic extraction + cache |
 | 6 | synthesize | `app/steps/synthesize.ts` | single-pass or EESV |
-| 7 | verify | `app/steps/verify.ts` | structural verify + repair |
-| 8 | state | `app/steps/state.ts` + `domain/yield-gate.ts` | state/open loops/delta + final yield proof |
+| 7 | verify | `app/steps/verify.ts` | structural verify + repair; high-risk outcomes require successful tool evidence |
+| 8 | state | `app/steps/state.ts` + `domain/yield-gate.ts` | state/open loops/resolved history/delta + final yield proof |
 | 9 | persist | `app/steps/persist.ts` | stage pending, apply compaction |
 | 10 | metrics | `app/steps/metrics.ts` | success / failure record |
 
@@ -120,8 +120,9 @@ pressure and deterministic extraction risk.
 Model routes are stage-specific but never inferred from mode. With no explicit
 configuration, Explore, Synthesize, and Verify all use the selected Pi model.
 `segmentationModel`, `summaryModel`, and `verificationModel` can independently
-override those routes. `prepareRun()` resolves credentials once per distinct
-provider/model route; call metrics preserve the actual route.
+override those routes. Credentials resolve lazily immediately before each
+stage's first network call and are reused for equivalent routes; call metrics
+preserve the actual route.
 
 ### Keep window and preprocessing
 
@@ -145,10 +146,11 @@ keeps chunked recovery instead of sending an oversized one-shot prompt to
 native summarization. Manual runs use the profile's absolute adaptive tail, so
 model-window size cannot turn an explicit command into a full-context no-op.
 
-Before summarization the pipeline serializes the full selected conversation,
-scrubs it, and writes that pre-prune backup; it then prunes redundant messages,
+Before summarization the pipeline serializes and scrubs the full selected
+conversation into an in-memory prepared backup, then prunes redundant messages,
 loads the previous verified summary plus bounded continuity state, checks the
-incremental extraction cache, and loads the project fingerprint.
+incremental extraction cache, and loads the project fingerprint. The backup is
+written atomically only after the matching native compaction is confirmed.
 
 ### Extract
 
@@ -187,16 +189,20 @@ deterministic fallback assembly when any budget or LLM call fails.
 ### Verify
 
 Primary: [`src/phases/verify.ts`](./src/phases/verify.ts). Scores the summary
-against the deterministic extraction. It checks for missing modified files,
-missing unresolved errors, missing high-confidence constraints, weak goal
-coverage, missing structure sections, suspicious fabricated file references,
-done/unresolved inconsistencies, missing explicit decisions, and missing
-open-loop coverage.
+against deterministic extraction, continuity, explicit focus/note steering,
+and source messages. It checks missing modified/read/deleted files, unresolved
+errors, high-confidence constraints, weak goal coverage, missing structure,
+suspicious fabricated paths, done/unresolved inconsistencies, explicit
+decisions, open loops, and unsupported high-risk outcome claims. Claims such as
+“tests passed” require matching source prose or a successful related tool
+result.
 
 **Repair order is intentional:** (1) deterministic patch first (free,
 idempotent) → (2) one LLM patch only in `thorough` mode if still insufficient
-→ (3) replace lower-scoring output with the deterministic quality floor → (4)
-reject unless final verification has no gaps and meets the verified threshold.
+→ (3) replace lower-scoring output with a deterministic quality floor built
+only from extraction, continuity, and steering → (4) reject unless final
+verification has no gaps and meets the verified threshold. Untrusted chunk
+prose is never an input to the quality floor.
 Final verification runs again after continuity injection. The final scalar is
 reported as repaired **verification coverage**, alongside the pre-repair source
 score and fallback provenance; it is not labeled as raw synthesis quality.
@@ -207,13 +213,13 @@ Markdown prefixes and multiline wrapping cannot create false missing-error gaps.
 
 ## EESV hardening and control surfaces
 
-- **Canonical summary IR** accepts recognized H1/H2/H3 headings, preserves Progress subsections, and merges duplicate canonical kinds before state mutation.
-- **Typed verification gaps** drive mandatory deterministic repair; collision-aware path needles prevent basename cross-satisfaction. Provenance is persisted and shown before optional approval.
+- **Canonical summary IR** accepts recognized H1/H2/H3 headings outside fenced code, preserves Progress subsections, and merges duplicate canonical kinds before state mutation.
+- **Typed verification gaps** drive mandatory deterministic repair; collision-aware path needles prevent basename cross-satisfaction. Tool/file provenance and normalized semantic evidence are indexed once per verification pass, and truncated or delimiter-incomplete LLM patches are rejected. Provenance is persisted and shown before optional approval.
 - **Fine tool semantics** separate read/search/list/mutate/delete/execute operations. Pruning deduplicates only identical idempotent access signatures.
-- **Unified token planning** uses a run-bound estimator with bounded process-shared provider/model calibration, counts structured tool-call arguments, preserves an adaptive recent tail, targets mode-specific post-compaction headroom, reserves 25% of the synthesis allowance for deterministic post-summary state sections, and reserves/reconciles every request against the mode's aggregate prompt/output-token caps.
-- **Security boundaries** scrub high-confidence secrets before provider calls and durable cache/backup/state writes; PII scrubbing is opt-in.
-- **Policy controls** include focus weighting, exact call/latency budgets, fail-closed manual approval, online damage monitoring, and persisted open-loop overrides.
-- **Release gate** (`bun run gate`) covers adversarial parser, verification, tool, cache, budget, scrub and damage fixtures.
+- **Unified token planning** uses a run-bound estimator with bounded process-shared provider/model calibration, counts structured tool-call arguments, preserves an adaptive recent tail, targets mode-specific post-compaction headroom, reserves bounded deterministic post-summary state sections, clamps every provider request to the model's advertised output limit, and reserves/reconciles every request against aggregate prompt/output-token caps. Missing provider usage is estimated conservatively. Tool exchanges remain atomic; oversized result bodies are head/tail bounded only for synthesis after full deterministic extraction.
+- **Security boundaries** recursively scrub structured messages before host serialization or provider calls, redact secret-bearing primitive values, and scrub plus hard-cap exploration tool feedback. PII scrubbing is opt-in. Backups remain unmaterialized until confirmed apply.
+- **Policy controls** include focus weighting, exact call/latency budgets, default fail-closed manual approval, online damage monitoring, and persisted open-loop overrides. Interactive review time is outside the pipeline deadline.
+- **Release gates** (`bun run gate`, `bun run bench`) cover adversarial parser, verification, tool, cache, budget, scrub and damage fixtures plus bounded p95 regressions for extraction, pruning, chunking, summary parsing, and path matching.
 
 ## State, caching & persistence
 
@@ -221,10 +227,12 @@ Post-verification, `app/steps/state.ts` + `src/utils/state.ts` enrich the
 summary, then `domain/yield-gate.ts` measures the final replacement. Planning
 has already reserved the bounded post-synthesis enrichment band by reducing the
 retained tail; missing the original target or 10% net-saving floor still throws
-before a `StatedRc` can reach staging/apply. `session_before_compact` only stages a passing candidate; after
-the host emits the matching `session_compact`, `app/steps/persist.ts` commits
-reusable state and success telemetry. Aborted/unconfirmed candidates write
-neither, and the UI reports `Applied` only after that correlated commit.
+before a `StatedRc` can reach staging/apply. `session_before_compact` only stages
+a passing candidate; after the host emits the matching `session_compact`,
+`app/steps/persist.ts` commits reusable state, the prepared conversation backup,
+and success telemetry. Aborted/unconfirmed candidates write none of them. The
+UI reports `Applied` only after that correlated commit and emits a separate
+warning if any durable side effect was partial.
 `ui/error-format.ts` converts verification/yield failures to one bounded,
 content-free diagnostic and collapses unknown multiline errors; full stacks are
 suppressed by default and emitted only under explicit `DEBUG=smart-compact`.
@@ -238,39 +246,44 @@ leaving one content-free safe-fallback notice in the UI.
 | Concern | Where | Notes |
 | --- | --- | --- |
 | Open-loop injection | `utils/state.ts` | inserted before Next Steps via the canonical parser |
-| `CompactionState` | `utils/state.ts` | conservatively merged and bounded across goal wording changes; newer file evidence resolves delete/present contradictions |
+| `CompactionState` | `utils/state.ts` | immutable project/session/branch-head snapshots; descendants resolve the newest matching ancestor and siblings never overwrite each other |
 | Continuity Ledger | `utils/state.ts` | prior facts carry forward until positive resolution evidence or an explicit override; goal shifts become non-destructive breadcrumbs |
 | Cross-compaction delta | `utils/state.ts` | "Changes Since Last Compaction" section |
-| Incremental extraction cache | `utils/cache.ts` + `utils/id-fingerprint.ts` | bounded SHA-256 prefix fingerprint + tail; safe only when the pruned prefix still matches |
-| Synthesis cache | `infra/synthesis-cache.ts` | behavior key includes normalized focus, route, mode, budgets, and reasoning |
-| Session-log recovery | `utils/session-log.ts` | streaming JSONL parse; bypasses pi-toolkit truncation by entry-id mapping |
-| Project fingerprint | `utils/fingerprint.ts` | locked read/merge/write; language/framework/key dirs stay bounded across sessions |
+| Incremental extraction cache | `utils/cache.ts` + `utils/id-fingerprint.ts` | bounded SHA-256 prefix fingerprint + tail; an exact pruned-payload match is reused directly, while incremental reuse is allowed only when the pruned prefix still matches |
+| Synthesis cache | `infra/synthesis-cache.ts` | behavior key includes normalized focus, route, mode, profile limits, run-level call/input/latency limits, and reasoning |
+| Session-log recovery | `utils/session-log.ts` | async bounded-memory JSONL scan with event-loop yields and a bounded path cache; bypasses pi-toolkit truncation by entry-id mapping without dropping late active-branch IDs at a fixed byte cap |
+| Project fingerprint | `utils/fingerprint.ts` | locked read/merge/write; language/framework/key dirs stay bounded and `sessionCount` tracks distinct hashed session identities |
 | Damage detection | `utils/damage.ts` | best-effort post-compaction regression signals |
 | Context graph | `infra/context-graph.ts` | SQLite FTS5 facts + file edges; 2,000 non-structural nodes per project |
 
 Apply-confirmed verified state is queued and duplicate updates coalesce only
-for the exact project/session/branch head, then indexed on the next event-loop
-turn so SQLite work is not part of the native
-compaction hook's latency. `infra/context-graph.ts` adapts the same synchronous
-query/transaction contract to `bun:sqlite` in Bun tests and `node:sqlite`
-`DatabaseSync` in Pi's Node runtime; the packed release audit exercises both.
-The referenced queue timer survives extension
+for the exact project/session/branch head. Replacing an existing key refreshes
+that pending value even at capacity; a divergent 65th key is rejected rather
+than evicting accepted work. The next event-loop turn drains the accepted
+batch through one reused SQLite connection, so connection setup and graph work
+are not part of the native compaction hook's latency.
+`infra/context-graph.ts` adapts the same fail-closed transaction contract to
+`bun:sqlite` in Bun tests and `node:sqlite` `DatabaseSync` in Pi's Node runtime;
+the packed release audit exercises both. The queue timer survives extension
 shutdown/reload in the process; graph data is derived and a later cumulative
-state safely supersedes a missed update after a hard process kill. Fact occurrences are branch-head scoped; state, recall, and resolution use the
+state safely supersedes a missed update after a hard process kill. Fact
+occurrences are branch-head scoped; state, recall, and resolution use the
 complete host-visible branch ancestry before equivalent facts are deduplicated.
 Schema v1 preserves user-confirmed manual memory but resets older derived
 compaction nodes once so sibling branches cannot inherit a last-writer identity.
-Recall starts from FTS5 lexical matches, expands one hop
-through file-reference edges, then applies session, branch, fact-kind,
-confidence, recency, and explicit-memory weights. Exact equivalent facts are
-deduplicated before bounded output. Resolved/superseded state is removed from
-the active FTS index; another project's rows are never eligible.
+Recall starts from FTS5 lexical matches whose rowids are the owning
+`context_nodes` rowids, expands one hop through file-reference edges, then
+applies session, branch, fact-kind, confidence, recency, and explicit-memory
+weights. Exact equivalent facts are deduplicated before bounded output.
+Resolved/superseded state is removed from the active FTS index; another
+project's rows are never eligible.
 
 **Important retention limits:** pending in-memory compaction 5 min · exploration
 tool-support cache 1 h / 128 routes · token calibration 128 routes · extraction
-cache 1 h · compaction state 7 d · context graph 2,000 non-structural fact nodes
-and 500 active manual memories per project · remediation hints 7 d · metrics and damage
-JSONL logs 5 MiB each.
+cache 1 h · compaction state 7 d / 64 snapshots · context graph 2,000
+non-structural fact nodes, 64 pending branch-head updates, and 500 active manual
+memories per project · remediation hints 7 d · metrics and damage JSONL logs
+5 MiB each · one exploration tool result 12,000 characters.
 
 ## Concurrency & safety model
 
@@ -288,30 +301,35 @@ loop). `consume()` returns a discriminated result:
 | `ok` | fresh payload for this session |
 | `empty` | nothing staged |
 | `expired` | older than the 5-minute TTL |
-| `mismatch` | staged by a **different** session (cross-session leak guard) |
+| `mismatch` | staged by a different session, project, or non-ancestor branch head |
 
 Session identity comes from [`infra/session-identity.ts`](./src/infra/session-identity.ts):
 a real id when the host exposes one, otherwise a per-call unforgeable
-`unresolved:<uuid>` — two unresolved sessions can never collide.
+`unresolved:<uuid>` — two unresolved sessions can never collide. Apply also
+requires the staged branch head to be the current head or one of its visible
+ancestors, so navigation to a sibling branch cannot consume stale payload.
 
 ### Cancellation deadlines
 
-Some providers ignore `AbortSignal`. The auto-trigger therefore uses a shared
-[`ExternalCancellation`](./src/app/run-smart-compact.ts) handle as a second line
-of defense: an outer `setTimeout` fires `abort()` and sets `timedOut`, and every
-side-effect gate in the orchestrator checks that flag before writing state or
-applying compaction. The caller waits for safe pipeline unwind; no unsafe
-`Promise.race` hard return can leave work running past the hook lifecycle.
+Automatic compaction combines the host event's `AbortSignal` with its own
+deadline through a shared [`ExternalCancellation`](./src/app/run-smart-compact.ts)
+handle. Either source calls `abort()`, and every side-effect gate in the
+orchestrator checks the shared state before writing or applying compaction.
+The caller waits for safe pipeline unwind; no unsafe `Promise.race` hard return
+can leave work running past the hook lifecycle.
 
 ### Filesystem & concurrency
 
 JSON/text cache writes use [`src/infra/fs.ts`](./src/infra/fs.ts): private
 artifact directories are 0700 and files are 0600; atomic temp-file + rename
 prevents half-truncated readers but intentionally does not claim fsync/power-loss
-durability. Append/trim operations use a `mkdir`-based cross-process lock and
-fail closed if ownership cannot be acquired, so sessions cannot interleave
-bytes. SQLite supplies its own WAL durability. Native continuity handoffs are
-one-shot, bounded, and keyed by project + session + branch head.
+durability. Append/trim operations run asynchronously, yield before synchronous
+filesystem work, and hold a `mkdir`-based cross-process lock for the complete
+transaction. Lock ownership is reclaimed by atomic rename, never by deleting a
+possibly renewed lease in place. Sessions therefore cannot interleave bytes or
+steal a live successor's lock. SQLite supplies its own WAL durability. Native
+continuity handoffs are one-shot, bounded, and keyed by project + session +
+branch head.
 
 ## Provider awareness
 
@@ -324,17 +342,19 @@ drives pipeline behavior:
 | --- | --- |
 | `maxOutputTokens` | caps synthesis / patch budgets |
 | `supportsTools` (`true \| false \| "probe"`) | exploration tool-call probing |
-| `concurrencyLimit` | batch synthesis wave scheduling |
+| `concurrencyLimit` | bounded batch-synthesis worker-pool width |
 | `cacheStrategy` | prompt-cache retention per call |
 | `timeoutMultiplier` | auto-trigger hard-timeout headroom |
 | `singlePassTokenMultiplier` | single-pass vs chunked threshold |
 | `tokenRatioEstimate` | token estimation; refined by per-(provider,model) **EMA calibration** |
 
-The Codex limiter is hybrid. Custom Codex endpoints receive
-`max_output_tokens` through Pi AI's payload hook. The ChatGPT subscription
-endpoint rejects every wire output-cap field, so it uses a derived 15–90s
-per-call stream watchdog plus a visible-output ceiling; aborts route to the
-phase's deterministic fallback.
+Every provider call is raced against one aborting hard deadline so a transport
+that ignores cancellation cannot keep the run lock indefinitely. Custom Codex
+endpoints also receive `max_output_tokens` through Pi AI's payload hook. The
+ChatGPT subscription endpoint rejects every wire output-cap field, so its
+deadline is derived from the requested output allowance (15–90s) and paired
+with a visible-output ceiling. Deadline failures route to the phase's
+deterministic fallback.
 
 ### Provider evaluation and routing evidence
 
@@ -414,14 +434,14 @@ The code is organized into six layers, each with a single responsibility.
 | `app/run-context.ts` | typed stage chain (`RcBase → … → StatedRc`) |
 | `app/mode-policy.ts` | Auto selector and finite Fast/Balanced/Thorough policies; legacy Aggressive maps to Fast |
 | `app/pending-slot.ts` | encapsulated pending-compaction state cell |
-| `app/steps/prepare.ts` | resolve config, auth, provider caps |
-| `app/steps/window.ts` | pick the prefix of messages to compact |
+| `app/steps/prepare.ts` | resolve config, provider caps, budgets, and cancellation; stage auth resolves lazily |
+| `app/steps/window.ts` | pick the prefix using provider-calibrated synthesis and deterministic post-processing bounds |
 | `app/steps/recover.ts` | recover full content for log-truncated messages |
 | `app/steps/tier.ts` | choose compaction tier (none / light / full) |
 | `app/steps/extract.ts` | pruning + deterministic extraction with incremental cache |
 | `app/steps/synthesize.ts` | single-pass / EESV synthesis |
-| `app/steps/verify.ts` | structural verification + repair |
-| `app/steps/state.ts` | enrich summary with state machine + open loops |
+| `app/steps/verify.ts` | structural verification + repair with tool-result trust boundaries |
+| `app/steps/state.ts` | enrich summary with state, open loops, and recent resolved-error history |
 | `app/steps/persist.ts` | apply compaction, save fingerprint, persist state |
 | `app/steps/metrics.ts` | record success / failure metrics |
 
@@ -450,14 +470,14 @@ All external-world interaction.
 
 | File | Responsibility |
 | --- | --- |
-| `infra/fs.ts` | atomic writes, advisory locks |
+| `infra/fs.ts` | atomic writes, advisory locks, and yielding async append/trim |
 | `infra/paths.ts` | canonical cache/session/backup paths |
 | `infra/git.ts` | cached git-root discovery |
 | `infra/clock.ts` | injectable wall clock |
 | `infra/llm-client.ts` | LLM seam, custom-Codex wire cap, and ChatGPT Codex stream watchdog |
 | `infra/services.ts` | per-run services container |
 | `infra/session-identity.ts` | robust session-id resolution with opaque `unresolved:` fallback |
-| `infra/ai-messages.ts` | boundary adapters between `LlmMessage` and pi-ai `Message` |
+| `infra/ai-messages.ts` | validated message upcasts and recursive pre-serialization redaction |
 
 ### Utility layer (`src/utils/`)
 

package/CHANGELOG.md

@@ -1,8 +1,51 @@
 # Changelog
 
-## [Unreleased]
+## [9.1.0] - 2026-08-11
 
-No changes yet.
+### Fixed
+
+- Provider-bound messages are recursively redacted before host serialization, secret-bearing object keys redact non-string primitive values, and exploration tool results are redacted and hard-capped before re-entering model context.
+- Native apply now requires matching project/session branch ancestry, automatic hooks forward host cancellation into the shared abort path, stale run-lock leases are reclaimed atomically, and retained state snapshots are capped.
+- Incremental extraction preserves bounded evidence and exact cache payloads; shell mutation tracking covers explicit literal targets without treating arbitrary command text as paths.
+- Canonical summary parsing ignores headings inside fenced code, verifier patches reject truncation and unclosed fences, file-operation provenance is indexed once, and exact done-file checks no longer accept path collisions.
+- Stable normalized goal identities prevent paraphrase-only goal shifts, and task-specific completion evidence can close completed follow-up loops without generic acknowledgements doing so.
+- Zero/invalid context-window metadata follows one safe percentage policy, metrics tail reads honor the actual byte count, and nested `provider/model` identifiers are accepted by command and tool routing.
+
+### Performance
+
+- Window planning and extraction reuse token, tool-call, and referenced-file indexes; exact pruned extraction cache hits bypass recomputation.
+- Batch synthesis uses a bounded worker pool, session-log path discovery is cached, metrics appends yield off the event loop, state ancestry lookup is filename-indexed, and pending-slot newest lookup is constant-time.
+- Context-graph FTS rowids match their content rows, queued same-branch updates refresh in place, SQLite transactions share fail-closed rollback semantics, and one connection is reused per drain.
+
+## [9.0.0] - 2026-08-09
+
+### Changed
+
+- Major hardening release: verified host-correlated apply, strict command/tool inputs, bounded automatic compaction, model-aware output ceilings, branch-scoped continuity, and fail-closed yield and evidence contracts are now enforced end to end.
+
+### Fixed
+
+- Command and tool inputs now share one strict parser. Control tokens are consumed only from the left edge, `--note`/`--` preserve literal notes, file paths are not mistaken for model IDs, and invalid modes or budgets fail explicitly.
+- Verified state is stored as immutable project/session/branch-head snapshots; sibling branches no longer overwrite continuity, loop overrides re-index the updated graph, and synthesis cache keys include every output-affecting run budget.
+- Deleted-file evidence now survives chunk synthesis, fallback assembly, verification, and repair. Modified/deleted paths count only in their canonical sections; unrelated parallel tool calls and substantive assistant constraints are no longer pruned.
+- Every provider transport has an aborting hard deadline, stale filesystem locks are reclaimed asynchronously, backup listing reads bounded headers, and session-log recovery scans large active logs incrementally without a 50 MiB correctness cutoff.
+- Result and approval UI expose generation fallbacks, source-versus-repaired coverage, remaining-gap counts, and the actual summary preview. Project fingerprints count distinct sessions instead of compaction runs.
+- Verification now grounds high-risk completion claims only against successful tool results or deterministic resolved-error/file evidence, removes unsupported assistant-only claims, and builds the quality floor only from deterministic extraction, continuity, and explicit steering.
+- Session-log recovery preserves exact entry/message identity; synthesis cache keys include recovered content and every provider route; full cache values are cloned without sharing nested mutable state.
+- Boundary indexes are integer-canonicalized, exploration arguments reject empty broad matches, extraction evidence is bounded with visible overflow counts, and delayed tool results remain paired under the chunk ceiling.
+- Missing provider usage is conservatively estimated for budgets and telemetry. Budget-driven synthesis fallbacks are recorded explicitly rather than appearing as provider-free success.
+- Scrubbed backups are prepared without I/O and written only after correlated native apply. Partial persistence is reported separately, stale state/temp artifacts are removed, and stage transitions now enforce runtime order and required fields.
+- Manual verified-summary review is enabled by default and excluded from the pipeline deadline; the preview wraps long evidence lines and dashboards hide unavailable quality scores.
+- Cross-process JSONL retention now holds the append/trim lock for the complete transaction, overflow planning accounts for host-only fixed context, and long command failures preserve a bounded error window from the middle of pruned tool output.
+- Single-batch synthesis fallbacks are visible and never cached as clean output; verification refreshes the run-wide LLM-call total; transient tool-probe failures no longer poison capability caches; profile overrides enforce integer and cross-field bounds.
+- Verified-summary review is one scrollable Apply/Cancel screen, unavailable preflight choices explain why Enter is blocked, exact backup payloads materialize only after native apply, and restore forks from the recorded branch leaf or opens an isolated session for legacy backups.
+- Unresolved host sessions stop before pipeline work, branch ancestry snapshots are bounded while retaining compaction state heads, Git-root caching is bounded and expiring, and context-graph directories are normalized to owner-only permissions.
+- Window planning now translates provider output caps through the calibrated local estimator and budgets bounded deterministic repair/state additions, preventing late target rejection after a high-saving run.
+- Secret scrubbing recognizes provider-specific Google/Stripe/GitLab/npm credentials, AWS secret assignments, credential-bearing object keys, and passwords in connection URIs.
+- Incremental extraction caches retain failed-operation signatures, so only the same later operation can resolve a cached error; recent resolved errors remain visible in the continuation summary.
+- Automatic compaction shows live phase progress and is bounded to 60 seconds and four LLM calls before native recovery. Context-graph queue saturation now rejects and reports the newest update instead of silently evicting accepted state.
+- CI actions are pinned to immutable commits, the redundant adversarial rerun was removed from CI, and restore tests delete their temporary HOME directories.
+- Every tracked provider request is clamped to the model's advertised output limit before budget reservation and dispatch. The release check now includes p95 hot-path regression limits and a host-correlated end-to-end auto-compaction/apply lifecycle test.
 
 ## [8.0.8] - 2026-08-09
 

package/README.md

@@ -38,12 +38,19 @@ Then run `/smart-compact` for an explainable preflight before anything changes.
 /smart-compact auto                                 # adaptive mode selection
 /smart-compact anthropic/claude-sonnet-4 fast      # explicit model + mode
 /smart-compact balanced --focus=auth                # preserve extra auth detail
+/smart-compact --note="keep balanced and fast terminology"
+/smart-compact -- preserve this note verbatim after the option boundary
 /smart-compact metrics                              # text metrics report
 /smart-compact dashboard                            # interactive dashboard
 /smart-compact restore                              # browse and restore backups
 /smart-compact loops                                # manage persisted open loops
 ```
 
+Command controls are consumed only from the left edge. Once note text starts,
+words such as `fast`, `balanced`, and paths such as `src/auth.ts` remain note
+content. Use `--note=...` or `--` when the boundary should be explicit. Invalid
+tool modes and budgets return an error instead of silently using defaults.
+
 At 60% context usage by default, the extension also participates in Pi's native
 compaction flow. Long-running agents can call `smart_compact`, `smart_recall`,
 and `smart_save_memory` directly.
@@ -68,7 +75,7 @@ The design principle is simple:
 
 > **Facts first. Synthesis second. Verification before apply.**
 
-Any unresolved verification gap rejects the custom summary before staging or apply. A zero-gap deterministic fallback is preferred over unverifiable model output; only failure of that fallback rejects the run. A successful compaction must also meet its mode target and at least 10% estimated net savings both before synthesis and after the final summary is measured. Automatic failures leave Pi free to use its native compactor, while manual failures leave the conversation unchanged.
+Any unresolved verification gap rejects the custom summary before staging or apply. High-risk outcome claims such as “tests passed” or “deployed” must match source messages or a successful related tool result; unsupported claims are removed deterministically. A zero-gap fallback built only from extraction, continuity, and explicit focus/note steering is preferred over unverifiable model output; untrusted chunk prose cannot become the quality floor. A successful compaction must also meet its mode target and at least 10% estimated net savings both before synthesis and after the final summary is measured. Automatic failures leave Pi free to use its native compactor, while manual failures leave the conversation unchanged.
 
 ## EESV pipeline
 
@@ -150,14 +157,18 @@ target, route, and boundary data stays under `D` instead of crowding the decisio
 - `D` toggles calibrated estimator, target, boundary, and route details.
 - `M` opens Advanced model selection and replans with that route's calibration.
 
-Values remain estimates until the next provider turn reports usage. Smart
-Compact therefore uses `~`/`≤` language, measures the completed summary again
-before staging, and reports final success only after Pi confirms the matching
-`session_compact` run ID. A single long user turn may be split at a safe message
-boundary: its older prefix is verified into the summary while the budgeted
-working tail stays raw. Tool exchanges are summarized or retained as complete
-call/result pairs, never split. If Verify, yield, provider, or native apply
-fails, the UI shows one bounded actionable line without evidence text or a
+Provider-reported token usage is used when available; missing or partial usage
+is conservatively estimated for both metrics and aggregate budgets. Every
+provider request is clamped to the selected model's advertised output limit
+before reservation and dispatch. Smart Compact uses `~`/`≤` language for
+post-compaction estimates, measures the completed summary again before staging,
+and reports final success only after Pi confirms the matching
+`session_compact` run ID. A single long user turn may be
+split at a safe message boundary: its older prefix is verified into the summary
+while the budgeted working tail stays raw. Tool exchanges remain complete
+call/result pairs; oversized result evidence is head/tail bounded only in the
+synthesis prompt after deterministic extraction has consumed the full input.
+If Verify, yield, provider, or native apply fails, the UI shows one bounded actionable line without evidence text or a
 JavaScript stack. A successful `100/100` is labeled **verification coverage**;
 the source score and deterministic/LLM/fallback provenance remain visible so
 repaired coverage is never presented as raw synthesis quality. Stack diagnostics
@@ -182,8 +193,8 @@ brief. `verbose` restores routine phase notices; full stack diagnostics require
   does **not** attempt unsupported non-contiguous compaction.
 - `--max-calls` accepts `1–100`.
 - `--max-input-tokens` accepts `10000–1000000` aggregate prompt tokens.
-- `--max-latency` accepts `5000–600000` milliseconds as a cancellation deadline; cancellation waits for safe pipeline unwind before returning.
-- Budget exhaustion degrades to deterministic summaries instead of dropping context.
+- `--max-latency` accepts `5000–600000` milliseconds as a provider/pipeline cancellation deadline; interactive summary review time is excluded.
+- Budget exhaustion is recorded as an explicit fallback outcome and degrades to deterministic summaries instead of dropping context.
 
 The tool exposes equivalent `focus`, `max_calls`, `max_input_tokens`, and
 `max_latency_ms` parameters.
@@ -206,7 +217,7 @@ the verified summary carries them forward. Automatic risk refinement may deepen
 analysis/repair strategy after extraction, but it does not mutate the profile
 allowance or retention window that was already used to prove the target.
 
-Output caps stop subsequent calls after observed usage reaches the threshold.
+Output caps stop subsequent calls after reported or conservatively estimated usage reaches the threshold.
 The ChatGPT Codex subscription endpoint rejects `max_output_tokens`,
 `max_tokens`, and `max_completion_tokens`; Smart Compact therefore enforces a
 15–90 second per-call watchdog plus a streamed visible-output ceiling and falls
@@ -286,11 +297,20 @@ guidance for reaching the target.
 - Tool-call/tool-result pair integrity at the compaction boundary
 - Collision-safe modified-file verification for monorepos
 - Bounded fixed-point repair for patchable verification gaps, followed by a zero-gap deterministic quality floor
+- High-risk success claims are grounded only in successful host/tool results or
+  deterministic resolved-error/file evidence; assistant prose is never proof
+  of its own claim.
+- The window planner converts provider output caps through the calibrated local
+  estimator and reserves bounded deterministic repair/state additions before
+  choosing the retained tail.
+- Recent resolved errors remain explicit in the Continuity Ledger instead of
+  disappearing when they leave the unresolved set.
 - Cross-session guard and five-minute TTL for pending summaries
 - Session-log recovery for older, truncated tool results
-- Full selected pre-prune conversation backups, scrubbed before a 0600 write;
-  marker-owned retention leaves foreign files in custom directories untouched
-- Private artifact directories are enforced as 0700 and files as 0600
+- The full selected pre-prune conversation is scrubbed and prepared in memory;
+  its 0600 backup is written only after the matching native compaction succeeds.
+  Marker-owned retention leaves foreign files in custom directories untouched.
+- Private artifact directories are enforced as 0700 and files as 0600; stale state snapshots and orphaned atomic-write temp files are removed during bounded retention sweeps.
 
 ### Secrets and PII
 
@@ -301,9 +321,10 @@ boundary:
 provider request · extraction cache · backup · state · context graph · pending summary
 ```
 
-It covers common API keys, cloud/GitHub/Slack tokens, JWTs, bearer tokens,
-private keys, and credential assignments. Optional email/phone/payment-card
-scrubbing is available through `scrubPii`.
+It covers common API keys (including Google and Stripe), AWS/GitHub/GitLab/npm/
+Slack tokens, JWTs, bearer tokens, private keys, secret-bearing object fields,
+generic credential assignments, and passwords embedded in connection URIs.
+Optional email/phone/payment-card scrubbing is available through `scrubPii`.
 
 Secret scrubbing is defense in depth, **not a replacement for proper secret
 handling or a dedicated DLP system**. See the
@@ -311,12 +332,13 @@ handling or a dedicated DLP system**. See the
 
 ### Approval and feedback
 
-- The manual preflight is the default decision point. `requireApproval: true`
-  additionally shows a fail-closed verified-summary **Apply / Cancel** review;
-  `false` avoids a redundant second modal. Fingerprint, continuity state,
-  context graph, and success telemetry commit only after the host confirms the
-  matching native `session_compact` event, and only then does the UI report
-  `Applied`.
+- Manual runs show a fail-closed verified-summary **Apply / Cancel** review by
+  default (`requireApproval: true`); set it to `false` only to opt out of the
+  second modal after preflight. Review time is not charged to the pipeline
+  deadline. Fingerprint, continuity state, context graph, prepared backup, and
+  success telemetry commit only after the host confirms the matching native
+  `session_compact` event. The UI reports `Applied` at that point and separately
+  warns if any durable persistence side effect was partial.
 - Online damage monitoring observes the first post-compaction messages and
   records re-read files or repeated context. Observations join the originating
   compaction by a local run id; missing evidence lowers coverage rather than
@@ -359,7 +381,7 @@ Add `smartCompact` to `~/.pi/agent/settings.json`:
     "backupEnabled": true,
     "scrubSecrets": true,
     "scrubPii": false,
-    "requireApproval": false,
+    "requireApproval": true,
     "maxLlmCalls": 8,
     "maxLlmInputTokens": 0,
     "codexMaxCallMs": 0,
@@ -426,13 +448,13 @@ the run fails closed before staging or apply.
 | `summaryThinkingLevel` | `minimal \| low \| medium \| high \| xhigh \| max \| null` | `minimal` | Reasoning level for synthesis and repair; provider default when null |
 | `segmentationThinkingLevel` | `minimal \| low \| medium \| high \| xhigh \| max \| null` | `minimal` | Reasoning level for exploration; provider default when null |
 | `autoTrigger` | `boolean` | `true` | Participate in Pi's native compact hook |
-| `autoTriggerTimeoutMs` | `number` | `120000` | Auto cancellation deadline; waits for safe pipeline unwind, never an unsafe hard return |
+| `autoTriggerTimeoutMs` | `number` | `120000` | Requested auto cancellation deadline; the synchronous hook clamps it to 60s and four LLM calls, shows live phase progress, then safely unwinds to native recovery |
 | `minContextPercent` | `number` | `60` | Auto/tool context gate; manual `/smart-compact` warns and bypasses it |
-| `backupEnabled` | `boolean` | `true` | Write a pre-compaction backup |
+| `backupEnabled` | `boolean` | `true` | Prepare a scrubbed pre-compaction backup; write it only after confirmed apply |
 | `backupDir` | `string` | `~/.pi/agent/compact-backups` | Empty config value uses this path |
 | `profiles` | object | built-ins | Per-profile numeric overrides |
 | `pinPaths` | `string[]` | `[]` | Always preserve matching paths |
-| `requireApproval` | `boolean` | `false` | Manual UI only; cancel/error fails closed |
+| `requireApproval` | `boolean` | `true` | Manual verified-summary review; cancel/error fails closed |
 | `scrubSecrets` | `boolean` | `true` | High-confidence credential redaction |
 | `scrubPii` | `boolean` | `false` | Email/phone/card-shaped redaction |
 | `maxLlmCalls` | integer `0–100` | `8` | Global ceiling combined with the selected mode |
@@ -546,8 +568,8 @@ coordinate hook order or prefer a single automatic compaction owner.
 
 ```bash
 bun install --frozen-lockfile
-bun run release:check   # typecheck + tests + adversarial gate + build + package audit
-bun run bench
+bun run release:check   # typecheck + tests + adversarial/performance gates + build + package audit
+bun run bench           # standalone hot-path p95 regression gate
 bun run compat:pi       # isolated latest-Pi compatibility check
 ```
 

package/SECURITY.md

@@ -6,7 +6,7 @@ Security fixes target the latest published version of `pi-smart-compact`.
 
 | Version | Supported |
 | --- | --- |
-| Latest `8.x` | ✅ |
+| Latest `9.x` | ✅ |
 | Older | ❌ |
 
 ## Reporting a vulnerability

package/dist/app/native-continuity-bridge.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"native-continuity-bridge.d.ts","sourceRoot":"","sources":["../../src/app/native-continuity-bridge.ts"],"names":[],"mappings":"AAUA,MAAM,WAAW,qBAAqB;IACpC,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;IAClB,YAAY,EAAE,MAAM,CAAC;CACtB;AASD,MAAM,WAAW,sBAAsB;IACrC,KAAK,CAAC,KAAK,EAAE,qBAAqB,EAAE,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;IACxD,IAAI,CAAC,KAAK,EAAE,qBAAqB,GAAG,MAAM,GAAG,IAAI,CAAC;IAClD,KAAK,CAAC,KAAK,CAAC,EAAE,qBAAqB,GAAG,IAAI,CAAC;IAC3C,IAAI,IAAI,MAAM,CAAC;CAChB;AAMD,wBAAgB,4BAA4B,CAAC,IAAI,GAAE;IACjD,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,GAAG,CAAC,EAAE,MAAM,MAAM,CAAC;IACnB,GAAG,CAAC,EAAE,MAAM,CAAC;CACT,GAAG,sBAAsB,CAqG9B"}
+{"version":3,"file":"native-continuity-bridge.d.ts","sourceRoot":"","sources":["../../src/app/native-continuity-bridge.ts"],"names":[],"mappings":"AAUA,MAAM,WAAW,qBAAqB;IACpC,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;IAClB,YAAY,EAAE,MAAM,CAAC;CACtB;AASD,MAAM,WAAW,sBAAsB;IACrC,KAAK,CAAC,KAAK,EAAE,qBAAqB,EAAE,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;IACxD,IAAI,CAAC,KAAK,EAAE,qBAAqB,GAAG,MAAM,GAAG,IAAI,CAAC;IAClD,KAAK,CAAC,KAAK,CAAC,EAAE,qBAAqB,GAAG,IAAI,CAAC;IAC3C,IAAI,IAAI,MAAM,CAAC;CAChB;AAeD,wBAAgB,4BAA4B,CAAC,IAAI,GAAE;IACjD,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,GAAG,CAAC,EAAE,MAAM,MAAM,CAAC;IACnB,GAAG,CAAC,EAAE,MAAM,CAAC;CACT,GAAG,sBAAsB,CA4G9B"}

package/dist/app/pending-slot.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"pending-slot.d.ts","sourceRoot":"","sources":["../../src/app/pending-slot.ts"],"names":[],"mappings":"AAAA,2EAA2E;AAE3E,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AACrD,OAAO,EAAoB,KAAK,sBAAsB,EAAE,MAAM,8BAA8B,CAAC;AAE7F,MAAM,MAAM,aAAa,GACrB;IAAE,IAAI,EAAE,IAAI,CAAC;IAAC,OAAO,EAAE,iBAAiB,CAAA;CAAE,GAC1C;IAAE,IAAI,EAAE,OAAO,CAAA;CAAE,GACjB;IAAE,IAAI,EAAE,SAAS,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,GAClC;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,CAAC;AAE3D,MAAM,WAAW,WAAW;IAC1B,GAAG,CAAC,OAAO,EAAE,iBAAiB,GAAG,IAAI,CAAC;IACtC,OAAO,CAAC,GAAG,EAAE,sBAAsB,GAAG,aAAa,CAAC;IACpD,KAAK,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAChC,SAAS,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC;IACvC,IAAI,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,QAAQ,CAAC,iBAAiB,CAAC,GAAG,IAAI,CAAC;IAC7D,IAAI,IAAI,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,kBAAkB;IACjC,KAAK,EAAE,MAAM,CAAC;IACd,GAAG,CAAC,EAAE,MAAM,MAAM,CAAC;IACnB,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB;AAOD,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,kBAAkB,GAAG,WAAW,CAoEvE"}
+{"version":3,"file":"pending-slot.d.ts","sourceRoot":"","sources":["../../src/app/pending-slot.ts"],"names":[],"mappings":"AAAA,2EAA2E;AAE3E,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AACrD,OAAO,EAAoB,KAAK,sBAAsB,EAAE,MAAM,8BAA8B,CAAC;AAE7F,MAAM,MAAM,aAAa,GACrB;IAAE,IAAI,EAAE,IAAI,CAAC;IAAC,OAAO,EAAE,iBAAiB,CAAA;CAAE,GAC1C;IAAE,IAAI,EAAE,OAAO,CAAA;CAAE,GACjB;IAAE,IAAI,EAAE,SAAS,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,GAClC;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,CAAC;AAE3D,MAAM,WAAW,WAAW;IAC1B,GAAG,CAAC,OAAO,EAAE,iBAAiB,GAAG,IAAI,CAAC;IACtC,OAAO,CAAC,GAAG,EAAE,sBAAsB,GAAG,aAAa,CAAC;IACpD,KAAK,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAChC,SAAS,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC;IACvC,IAAI,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,QAAQ,CAAC,iBAAiB,CAAC,GAAG,IAAI,CAAC;IAC7D,IAAI,IAAI,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,kBAAkB;IACjC,KAAK,EAAE,MAAM,CAAC;IACd,GAAG,CAAC,EAAE,MAAM,MAAM,CAAC;IACnB,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB;AAOD,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,kBAAkB,GAAG,WAAW,CAsFvE"}

package/dist/app/preflight.d.ts

@@ -1,12 +1,13 @@
 import type { ExtensionContext } from "@earendil-works/pi-coding-agent";
 import type { Api, Model } from "@earendil-works/pi-ai";
-import type { CompactConfig, EffectiveCompactionMode, ProfileConfig, SessionMessageEntry } from "../types.ts";
+import type { CompactConfig, EffectiveCompactionMode, ProfileConfig, ProviderCapabilities, SessionMessageEntry } from "../types.ts";
 import { type TokenCalibrationStore, type TokenEstimator } from "../utils/tokens.ts";
 import type { CompactionWindowPlan } from "./run-context.ts";
 export declare function preflightDamageMedian(cwd: string, config: CompactConfig): number;
 export interface PreparedPreflightProfile {
     profileCfg: ProfileConfig;
     estimator: TokenEstimator;
+    providerCaps: ProviderCapabilities;
     adapted: boolean;
     damageMedian: number;
 }

package/dist/app/preflight.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"preflight.d.ts","sourceRoot":"","sources":["../../src/app/preflight.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,iCAAiC,CAAC;AACxE,OAAO,KAAK,EAAE,GAAG,EAAE,KAAK,EAAE,MAAM,uBAAuB,CAAC;AAExD,OAAO,KAAK,EAAE,aAAa,EAAE,uBAAuB,EAAc,aAAa,EAAE,mBAAmB,EAAE,MAAM,aAAa,CAAC;AAI1H,OAAO,EAAsB,KAAK,qBAAqB,EAAE,KAAK,cAAc,EAAE,MAAM,oBAAoB,CAAC;AAEzG,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,kBAAkB,CAAC;AAG7D,wBAAgB,qBAAqB,CAAC,GAAG,EAAE,MAAM,EAAE,MAAM,EAAE,aAAa,GAAG,MAAM,CAMhF;AAED,MAAM,WAAW,wBAAwB;IACvC,UAAU,EAAE,aAAa,CAAC;IAC1B,SAAS,EAAE,cAAc,CAAC;IAC1B,OAAO,EAAE,OAAO,CAAC;IACjB,YAAY,EAAE,MAAM,CAAC;CACtB;AAED,kFAAkF;AAClF,wBAAgB,uBAAuB,CAAC,KAAK,EAAE;IAC7C,GAAG,EAAE,MAAM,CAAC;IACZ,YAAY,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC;IACzB,IAAI,EAAE,uBAAuB,CAAC;IAC9B,gBAAgB,EAAE,qBAAqB,CAAC;IACxC,MAAM,EAAE,aAAa,CAAC;IACtB,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB,GAAG,wBAAwB,CAkB3B;AAED,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,uBAAuB,CAAC;IAC9B,IAAI,EAAE,oBAAoB,GAAG,IAAI,CAAC;IAClC,MAAM,EAAE,oBAAoB,CAAC,QAAQ,CAAC,GAAG,qBAAqB,CAAC;IAC/D,UAAU,EAAE,aAAa,CAAC;IAC1B,WAAW,EAAE,MAAM,CAAC;IACpB,yBAAyB,EAAE,MAAM,CAAC;IAClC,cAAc,EAAE,MAAM,CAAC;IACvB,OAAO,EAAE,OAAO,CAAC;IACjB,YAAY,EAAE,MAAM,CAAC;IACrB,mBAAmB,EAAE,MAAM,CAAC;IAC5B,cAAc,EAAE,MAAM,CAAC;IACvB,WAAW,EAAE,MAAM,CAAC;IACpB,iBAAiB,EAAE,OAAO,CAAC;CAC5B;AAED,MAAM,WAAW,sBAAsB;IACrC,MAAM,EAAE,OAAO,EAAE,CAAC;IAClB,IAAI,EAAE,mBAAmB,EAAE,CAAC;IAC5B,aAAa,EAAE,MAAM,EAAE,CAAC;IACxB,WAAW,EAAE,MAAM,CAAC;IACpB,yBAAyB,EAAE,MAAM,CAAC;IAClC,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,mBAAmB,EAAE,MAAM,CAAC;IAC5B,cAAc,EAAE,MAAM,CAAC;IACvB,WAAW,EAAE,MAAM,CAAC;IACpB,iBAAiB,EAAE,OAAO,CAAC;CAC5B;AAED,uEAAuE;AACvE,wBAAgB,6BAA6B,CAC3C,GAAG,EAAE,gBAAgB,EACrB,YAAY,EAAE,KAAK,CAAC,GAAG,CAAC,EACxB,gBAAgB,EAAE,qBAAqB,GACtC,sBAAsB,CAqBxB;AAED,wBAAgB,mBAAmB,CACjC,GAAG,EAAE,gBAAgB,EACrB,YAAY,EAAE,KAAK,CAAC,GAAG,CAAC,EACxB,IAAI,EAAE,uBAAuB,EAC7B,gBAAgB,EAAE,qBAAqB,EACvC,MAAM,EAAE,aAAa,EACrB,YAAY,CAAC,EAAE,MAAM,EACrB,MAAM,GAAE,sBAA2F,GAClG,eAAe,CAoCjB"}
+{"version":3,"file":"preflight.d.ts","sourceRoot":"","sources":["../../src/app/preflight.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,iCAAiC,CAAC;AACxE,OAAO,KAAK,EAAE,GAAG,EAAE,KAAK,EAAE,MAAM,uBAAuB,CAAC;AAExD,OAAO,KAAK,EAAE,aAAa,EAAE,uBAAuB,EAAc,aAAa,EAAE,oBAAoB,EAAE,mBAAmB,EAAE,MAAM,aAAa,CAAC;AAIhJ,OAAO,EAA2D,KAAK,qBAAqB,EAAE,KAAK,cAAc,EAAE,MAAM,oBAAoB,CAAC;AAE9I,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,kBAAkB,CAAC;AAG7D,wBAAgB,qBAAqB,CAAC,GAAG,EAAE,MAAM,EAAE,MAAM,EAAE,aAAa,GAAG,MAAM,CAMhF;AAED,MAAM,WAAW,wBAAwB;IACvC,UAAU,EAAE,aAAa,CAAC;IAC1B,SAAS,EAAE,cAAc,CAAC;IAC1B,YAAY,EAAE,oBAAoB,CAAC;IACnC,OAAO,EAAE,OAAO,CAAC;IACjB,YAAY,EAAE,MAAM,CAAC;CACtB;AAED,kFAAkF;AAClF,wBAAgB,uBAAuB,CAAC,KAAK,EAAE;IAC7C,GAAG,EAAE,MAAM,CAAC;IACZ,YAAY,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC;IACzB,IAAI,EAAE,uBAAuB,CAAC;IAC9B,gBAAgB,EAAE,qBAAqB,CAAC;IACxC,MAAM,EAAE,aAAa,CAAC;IACtB,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB,GAAG,wBAAwB,CAmB3B;AAED,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,uBAAuB,CAAC;IAC9B,IAAI,EAAE,oBAAoB,GAAG,IAAI,CAAC;IAClC,MAAM,EAAE,oBAAoB,CAAC,QAAQ,CAAC,GAAG,qBAAqB,CAAC;IAC/D,UAAU,EAAE,aAAa,CAAC;IAC1B,WAAW,EAAE,MAAM,CAAC;IACpB,yBAAyB,EAAE,MAAM,CAAC;IAClC,cAAc,EAAE,MAAM,CAAC;IACvB,OAAO,EAAE,OAAO,CAAC;IACjB,YAAY,EAAE,MAAM,CAAC;IACrB,mBAAmB,EAAE,MAAM,CAAC;IAC5B,cAAc,EAAE,MAAM,CAAC;IACvB,WAAW,EAAE,MAAM,CAAC;IACpB,iBAAiB,EAAE,OAAO,CAAC;CAC5B;AAED,MAAM,WAAW,sBAAsB;IACrC,MAAM,EAAE,OAAO,EAAE,CAAC;IAClB,IAAI,EAAE,mBAAmB,EAAE,CAAC;IAC5B,aAAa,EAAE,MAAM,EAAE,CAAC;IACxB,WAAW,EAAE,MAAM,CAAC;IACpB,yBAAyB,EAAE,MAAM,CAAC;IAClC,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,mBAAmB,EAAE,MAAM,CAAC;IAC5B,cAAc,EAAE,MAAM,CAAC;IACvB,WAAW,EAAE,MAAM,CAAC;IACpB,iBAAiB,EAAE,OAAO,CAAC;CAC5B;AAED,uEAAuE;AACvE,wBAAgB,6BAA6B,CAC3C,GAAG,EAAE,gBAAgB,EACrB,YAAY,EAAE,KAAK,CAAC,GAAG,CAAC,EACxB,gBAAgB,EAAE,qBAAqB,GACtC,sBAAsB,CAsBxB;AAED,wBAAgB,mBAAmB,CACjC,GAAG,EAAE,gBAAgB,EACrB,YAAY,EAAE,KAAK,CAAC,GAAG,CAAC,EACxB,IAAI,EAAE,uBAAuB,EAC7B,gBAAgB,EAAE,qBAAqB,EACvC,MAAM,EAAE,aAAa,EACrB,YAAY,CAAC,EAAE,MAAM,EACrB,MAAM,GAAE,sBAA2F,GAClG,eAAe,CAqCjB"}