Complete rewrite

[BadLiveware]

Note

High Risk
High risk because it replaces the extension’s compaction path with a new model-assisted MRC flow, adds hidden-reference state/lookup tooling, and introduces optional external LLM API calls that affect session data handling and runtime behavior.

Overview
Switches the extension from pi-vcc to pi-mrc (Model-Reference Compactor), updating naming, package metadata, and user docs to center on KEEP/REF/DROP chunk classification, exact handle lookup via mrc_lookup, and cache-stable prompt shaping.

Adds a new MRC runtime surface: compaction interception controls (/pi-mrc, /pi-mrc-off, /pi-mrc-on), compaction report capture + /pi-mrc-report history/artifacts, and /pi-mrc-dump-context for exporting structured context (including optional raw context/provider payload dumps). The before-compact hook now builds an MRC cut, filters internal ref/report messages, produces an MRC report, and records a latest-compaction ref index for lookup/continuity.

Introduces a benchmark + Docker workflow: new offline compaction runner, synthetic/real-session fixtures, cache-boundary gates, and scripts to run/compare compactors (including a model-reference-selector compactor that can use a real OpenAI-compatible classifier or a mock).

^{Reviewed by Cursor Bugbot for commit d6bf9db. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cubic-dev-ai]

30 issues found across 55 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/extract/goals.ts">

<violation number="1" location="src/extract/goals.ts:14">
P2: `PREFERENCE_WITH_TASK_RE` is a near-duplicate of `TASK_RE`, differing only by the omission of "test". If unintentional, "don't use X, test with Y" will be wrongly filtered. Either reference `TASK_RE` directly (if semantics should match) or add a comment explaining the deliberate difference.</violation>

<violation number="2" location="src/extract/goals.ts:46">
P2: `isPreferenceOnly` unconditionally returns `true` when `DIRECT_PREFERENCE_RE` matches, even if the text also contains a task verb. This causes lines like "please use X to fix Y" or "I prefer to implement Z" to be incorrectly filtered out as preference-only, losing legitimate goals.</violation>
</file>

<file name="src/commands/pi-vcc-report.ts">

<violation number="1" location="src/commands/pi-vcc-report.ts:32">
P2: Empty `catch {}` silently swallows errors from `getEntries()`. Consider at minimum adding a comment explaining why errors are expected, or logging at debug level so failures are diagnosable.</violation>
</file>

<file name="src/commands/pi-vcc-strategy.ts">

<violation number="1" location="src/commands/pi-vcc-strategy.ts:20">
P2: The function name `registerPiVccMrCommand` is misleading — it registers three commands (`pi-vcc-mr`, `pi-vcc-pv`, `pi-vcc-off`), not just the model-reference one. Consider renaming to something like `registerPiVccStrategyCommands`.</violation>
</file>

<file name="scripts/bench-compaction.ts">

<violation number="1" location="scripts/bench-compaction.ts:21">
P2: Validate that `realLimitRaw` parses to a valid integer. `Number.parseInt` returns `NaN` for non-numeric input, and passing `NaN` as a limit could silently produce empty results or unexpected behavior downstream.</violation>
</file>

<file name="bench/compaction/real-sessions.ts">

<violation number="1" location="bench/compaction/real-sessions.ts:15">
P3: Use `join` from `node:path` instead of manual string concatenation for constructing file paths. The module is already imported and `join` handles trailing-slash normalization automatically.</violation>
</file>

<file name="src/tools/lookup.ts">

<violation number="1" location="src/tools/lookup.ts:85">
P2: The `list` parameter is defined in the schema but never read in `execute`. Listing happens as a fallthrough when neither `ref` nor `query` is set, so `list` has no effect on behavior. If the intent is to let `list: true` force listing mode (even alongside other params), add an explicit check; otherwise remove the unused parameter to avoid misleading callers.</violation>
</file>

<file name="README.md">

<violation number="1" location="README.md:141">
P2: The features list still claims "5 semantic sections" but the updated sections table now documents 8+ distinct sections (adding Evidence Handles, Current Scope, and multiple Recent\* sections). Update the feature bullet to reflect the actual section count.</violation>
</file>

<file name="bench/compaction/offline-runner.ts">

<violation number="1" location="bench/compaction/offline-runner.ts:273">
P2: Duplicate function: `leakProbe` is identical to `termProbe`. Consider reusing `termProbe` directly where `leakProbe` is called, or differentiate the semantics if leak detection should behave differently (e.g., checking presence regardless of source applicability).</violation>
</file>

<file name="src/strategies/model-reference.ts">

<violation number="1" location="src/strategies/model-reference.ts:66">
P2: The empty `catch {}` silently swallows all errors when reading the auth file (missing file, permission denied, malformed JSON). Consider logging at debug level so users can diagnose why the real classifier isn't being used.</violation>

<violation number="2" location="src/strategies/model-reference.ts:97">
P2: The `classifierMs` stat actually measures the entire `compactWithModelReference` function execution time (normalization, filtering, chunking, classification, and rendering), not just the classifier latency. Rename to `totalMs` or measure only the classification step separately.</violation>
</file>

<file name="bench/compaction/synthetic-cases.ts">

<violation number="1" location="bench/compaction/synthetic-cases.ts:128">
P2: Duplicate `recallTerms` property in object literal — the second silently overwrites the first. If both are intentionally identical, remove the duplicate.</violation>
</file>

<file name="src/core/classifier.ts">

<violation number="1" location="src/core/classifier.ts:150">
P1: Parsing bug: the first line after the SUBGOALS section is always skipped. When a non-subgoal line (e.g., `KEEP:`) is encountered while `inSubgoals` is true, the code sets `inSubgoals = false` but still executes `continue`, so that line is never processed by subsequent matchers.</violation>
</file>

<file name="src/core/chunk-model.ts">

<violation number="1" location="src/core/chunk-model.ts:28">
P3: The doc comment example `"goal:0"` is misleading — the actual ID generated for goals would be `"sessionGoal:0"` (since the `section` parameter is `"sessionGoal"`). Update the example to match the real IDs.</violation>

<violation number="2" location="src/core/chunk-model.ts:128">
P2: `RefIndex.entries` uses an inline type that duplicates `RefIndexEntry`. Use the named interface to avoid drift between the two definitions.</violation>
</file>

<file name="src/core/mock-classifier.ts">

<violation number="1" location="src/core/mock-classifier.ts:73">
P2: Operator precedence bug: `&&` binds tighter than `||`, so the `text.length < 120` guard only applies to the `text.includes("/")` branch. The regex alone (which matches nearly any text with a dot, like "e.g", "v1.0") will unconditionally return `SCORE.FILE_PATH`. Add parentheses to clarify or gate both branches with the length check.</violation>
</file>

<file name="src/commands/pi-vcc-dump-context.ts">

<violation number="1" location="src/commands/pi-vcc-dump-context.ts:39">
P2: ``includes("--raw")`` also matches `--raw-context`. Consider using a word-boundary check (e.g., `/\b--raw\b/` or splitting args) to avoid the false positive. Currently safe only because `isRawContext` is checked first, but this coupling is fragile.</violation>

<violation number="2" location="src/commands/pi-vcc-dump-context.ts:107">
P2: The `--raw` check is placed after context extraction, so if extraction fails the function returns early with an error before reaching the `--raw` handler. Since `dumpRawSessionJsonl` reads the session file directly and doesn't need `extracted`, move the `--raw` check before context extraction (similar to how `--raw-context` is handled).</violation>
</file>

<file name="src/core/dump-context.ts">

<violation number="1" location="src/core/dump-context.ts:311">
P2: Unused `key` variable — deduplication is case-sensitive here unlike the identical pattern used for decisions above. `key` should be used for case-insensitive deduplication, or a `Set` should be employed for O(1) lookups (matching the `seenDecisions` pattern).</violation>
</file>

<file name="src/core/compaction-report-history.ts">

<violation number="1" location="src/core/compaction-report-history.ts:27">
P2: Type guard `isPiVccCompactionReport` validates only 5 of ~15 required fields on `PiVccCompactionReport`. If a partially-formed object satisfies the minimal checks (has `compactor`, `version`, `sections`, `sourceMessageCount`, `tokensBefore`), downstream code in `recordKeyOf` will read `keptMessageCount` and `summaryChars` as `undefined`, causing silent deduplication inconsistencies. Consider validating at least the fields used in `recordKeyOf`.</violation>
</file>

<file name="src/hooks/before-compact.ts">

<violation number="1" location="src/hooks/before-compact.ts:262">
P2: Second `getSessionStrategy() === "off"` check contradicts the earlier "always handle explicit /pi-vcc marker" logic. When a user explicitly runs `/pi-vcc` with session strategy "off", the function passes the first guard, computes stats, then silently bails here—giving no feedback to the user about why their command didn't work. Either this check should also be guarded with `!isPiVcc`, or the first check's comment is misleading.</violation>

<violation number="2" location="src/hooks/before-compact.ts:284">
P1: Model-reference strategy path returns `details` without `compactor: "pi-vcc"` and never sets `pendingReport`, so the compaction report message is silently lost. The `session_compact` handler can't extract the report from either `event.compactionEntry.details` (no `compactor` field) or `pendingReport` (null). Also `lastCompactWasPiVcc` is not set, breaking toast suppression for explicit `/pi-vcc` invocations.</violation>
</file>

<file name="src/core/model-reference-stitch.ts">

<violation number="1" location="src/core/model-reference-stitch.ts:41">
P3: `KNOWN_KINDS` includes `"recall"` but `KIND_ORDER` does not define a sort position for it. If recall chunks ever appear in keep sections, they'll sort unpredictably at position 99. Consider adding an explicit entry or a comment explaining the omission.</violation>

<violation number="2" location="src/core/model-reference-stitch.ts:200">
P2: Convert `keepIds` to a `Set` before filtering to avoid O(n×m) lookups. `bundledIds` already uses a Set on the previous line — apply the same pattern here.</violation>
</file>

<file name="src/core/compaction-state.ts">

<violation number="1" location="src/core/compaction-state.ts:215">
P3: Redundant filter: when `includeRecallNote` is false, no recall layer exists in `layers`, so the `.filter(...)` branch is always a no-op. This makes the intent unclear to a reader. Consider simplifying to `const bodyLayers = layers;` (the recall layer is only ever added when `includeRecallNote` is true).</violation>
</file>

<file name="scripts/compare-compaction-refs.mjs">

<violation number="1" location="scripts/compare-compaction-refs.mjs:213">
P2: Use `!= null` (loose) instead of `!== null` (strict) to also exclude `undefined` values. When `stablePrefixTokens` is absent from a JSONL row, `undefined !== null` passes the filter, producing `NaN` deltas that corrupt the sort and report output.</violation>

<violation number="2" location="scripts/compare-compaction-refs.mjs:317">
P2: Docker images built during the run (`pi-vcc-bench-baseline-*`, `pi-vcc-bench-head-*`) are never removed. Consider adding `docker rmi` calls in the `finally` block (or gated behind `!keepWorktrees`) to avoid accumulating stale images on repeated runs.</violation>
</file>

<file name="src/core/build-sections.ts">

<violation number="1" location="src/core/build-sections.ts:101">
P2: Single `pendingReadPath` variable doesn't handle parallel tool calls correctly. When multiple read tool_calls appear consecutively (common with parallel tool use), each overwrites the previous path. This causes the first result to be attributed to the wrong file and subsequent results to be skipped entirely.

Consider using a queue/map keyed by tool call index, or pairing tool_calls with tool_results by `sourceIndex`.</violation>
</file>

<file name="src/core/compaction-report.ts">

<violation number="1" location="src/core/compaction-report.ts:77">
P2: These hardcoded section-name sets can silently fall out of sync with `CURRENT_SECTION_ORDER`. Consider deriving them from `CURRENT_SECTION_ORDER` or adding a compile-time/test assertion that their union equals `CURRENT_SECTION_ORDER` to prevent silent misclassification of new sections.</violation>

<violation number="2" location="src/core/compaction-report.ts:199">
P3: Duplicated limit-lookup logic between `capOf` and the `limit` field in `buildCompactionReport`. Extract a shared `limitOf(title)` helper so the two cannot diverge.</violation>
</file>

_{Partial review: This PR has more than 50 files, so cubic reviewed the highest-priority files first. During the trial, paid plans get a higher file limit.
You can try an ultrareview to bypass the file limit, comment @cubic-dev-ai ultrareview. Learn more.
Shadow auto-approve: would not auto-approve because issues were found.
Fix all with cubic}

[cubic-dev-ai]

5 issues found across 31 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="package.json">

<violation number="1" location="package.json:15">
P2: The npm package scope (`@sting8k`) doesn't match the GitHub repository owner (`BadLiveware`). This inconsistency makes it harder for consumers to verify package provenance and locate the canonical source.</violation>
</file>

<file name="src/core/model-reference-stitch.ts">

<violation number="1" location="src/core/model-reference-stitch.ts:5">
P2: The recall note no longer mentions `bundle:*` handles, though `renderRetrievableIndex` in this file still emits `bundle:*` entries. Consider adding a bullet or expanding the `ref:*` bullet to also cover `bundle:*` handle lookup, so the model knows both handle types are resolvable via `mrc_lookup`.</violation>
</file>

<file name="src/core/format.ts">

<violation number="1" location="src/core/format.ts:23">
P2: This `RECALL_NOTE` is appended to all compaction summaries (including the default/pv strategy path via `renderCompactionState(..., { includeRecallNote: true })` in `summarize.ts`). However, its new text references MRC-specific concepts ("MRC handles", "pi-mrc") that only exist when the model-reference strategy is active. Non-MRC sessions will receive a confusing instruction about handles that aren't present in their summary.</violation>
</file>

<file name="src/hooks/before-compact.ts">

<violation number="1" location="src/hooks/before-compact.ts:208">
P2: Anchor messages (`PI_MRC_ANCHOR_TYPE`) are collected by `messageFromEntry` but not filtered by `isInternalMessage`, so they leak into the summarizer input. Import and include `isMrcAnchorMessage` in the filter to match how report and reference messages are already excluded.</violation>
</file>

<file name="src/core/mrc-reference-journal.ts">

<violation number="1" location="src/core/mrc-reference-journal.ts:216">
P2: Missing per-ref validation in `refsFromMrcReferenceEntries` — unlike `refsFromCompactionDetails` which filters `ref?.id && ref?.text`, this function pushes all entries from the array without checking required fields. Add a similar filter to guard against malformed stored data.</violation>
</file>

_{Shadow auto-approve: would not auto-approve because issues were found.
Fix all with cubic}

[cubic-dev-ai]

5 issues found across 21 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="scripts/bench-compaction.ts">

<violation number="1" location="scripts/bench-compaction.ts:22">
P2: `--real-limit` validation is bypassed for malformed numeric strings because `parseInt` truncates input (for example `1.5` or `10abc`).</violation>
</file>

<file name="src/core/classifier.ts">

<violation number="1" location="src/core/classifier.ts:175">
P2: Unconditionally ending `SUBGOALS` on the first non-matching line can silently discard later valid subgoal entries after one malformed status line.</violation>
</file>

<file name="src/core/build-sections.ts">

<violation number="1" location="src/core/build-sections.ts:111">
P2: Failed read results are not dequeued, which desynchronizes `pendingReadPaths` and can pair later read outputs with the wrong file path.</violation>
</file>

<file name="src/hooks/before-compact.ts">

<violation number="1" location="src/hooks/before-compact.ts:185">
P3: The report reason text mislabels `totalMs` as classifier/heuristic time, even though it is end-to-end compaction time.</violation>
</file>

<file name="src/tools/lookup.ts">

<violation number="1" location="src/tools/lookup.ts:83">
P2: `list` currently takes precedence over `ref`, so requests containing both fields skip exact handle lookup. Prefer handling `ref` when present (or explicitly gate list to ref-absent calls).</violation>
</file>

_{Tip: Review your code locally with the cubic CLI to iterate faster.
Shadow auto-approve: would not auto-approve because issues were found.
Fix all with cubic}

[cubic-dev-ai]

0 issues found across 10 files (changes from recent commits).

_{Shadow auto-approve: would not auto-approve. Auto-approval blocked by 9 unresolved issues from previous reviews.}

[cubic-dev-ai]

0 issues found across 2 files (changes from recent commits).

_{Shadow auto-approve: would auto-approve.}

[cubic-dev-ai]

1 issue found across 5 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="scripts/compare-compaction-refs.mjs">

<violation number="1" location="scripts/compare-compaction-refs.mjs:91">
P2: Cache failure counting no longer flags missing `firstChangedPromptLayer`, diverging from the benchmark’s cache gate definition and underreporting regressions.</violation>
</file>

_{Tip: Review your code locally with the cubic CLI to iterate faster.
Shadow auto-approve: would not auto-approve because issues were found.
Fix all with cubic}

[cubic-dev-ai]

2 issues found across 4 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="index.ts">

<violation number="1" location="index.ts:28">
P1: Provider request payloads are being buffered to disk unconditionally, bypassing the existing debug opt-in and increasing privacy risk for sensitive prompt/tool data.</violation>
</file>

<file name="src/core/context-buffer.ts">

<violation number="1" location="src/core/context-buffer.ts:97">
P1: Redact sensitive fields before persisting provider request payloads; this currently stores full request data in plaintext under `/tmp`.</violation>
</file>

_{Tip: Review your code locally with the cubic CLI to iterate faster.
Shadow auto-approve: would not auto-approve because issues were found.
Fix all with cubic}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 6fdb77b. Configure here.}

[cubic-dev-ai]

1 issue found across 3 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/core/model-reference-stitch.ts">

<violation number="1" location="src/core/model-reference-stitch.ts:131">
P2: Completed subgoal bullets reuse the `- <id> —` shape reserved for KEEP chunks, which can cause prior-summary KEEP parsers to misclassify subgoal labels as chunk IDs.</violation>
</file>

_{Tip: Review your code locally with the cubic CLI to iterate faster.
Shadow auto-approve: would not auto-approve because issues were found.
Fix all with cubic}

[cubic-dev-ai]

0 issues found across 6 files (changes from recent commits).

_{Shadow auto-approve: would auto-approve.}