Skip to content

feat: train named nested memory files in monorepos - #159

Merged
kunchenguid merged 17 commits into
kunchenguid:mainfrom
Athleticsmind:fm/backpass-nested-memory-v2
Sep 27, 2026
Merged

kunchenguid merged 17 commits into
kunchenguid:mainfrom
Athleticsmind:fm/backpass-nested-memory-v2

Conversation

@Athleticsmind

@Athleticsmind Athleticsmind commented Sep 27, 2026 •

Copy link
Copy Markdown
Contributor

Intent

sì, prepara la PR su backpass

Context needed to read the ask: the PR implements upstream issue #158, "Support nested AGENTS.md in monorepos (train per-directory memory files from the sessions that worked there)".
The issue: in a monorepo, memory is layered (root AGENTS.md plus nested ones such as apps/api/AGENTS.md that harnesses load only when working under that directory). backpass trains one file per repo: resolveMemoryFiles (src/memory.js) takes the first existing memoryFiles entry as primary and reports every other as separate (never updated, with a consolidate warning), correct for the AGENTS.md/CLAUDE.md pair but it means nested files are never trained. --target/--memory-file can name one nested file, but the evidence is still every session in the repo, so the gradient is noisy and root-only training pushes app-specific lessons into the always-loaded root. The issue proposes: nested memory files as additional weights (e.g. a nestedMemoryFiles list), keeping the AGENTS.md/CLAUDE.md pointer logic per directory; attributing each transcript to the directories it worked in via tool-call file paths and cwd, feeding a nested file only evidence from sessions that touched its subtree while root keeps cross-cutting evidence; routing each proposed edit to the most specific file owning the evidence; a per-file budget with an optional smaller default for nested files; one proposal and one apply review surface covering every file, each edit labeled by target file; two-session verbatim evidence, the token budget, and apply as the only writer unchanged.
The maintainer's triage comment on the issue (labeled ready-for-pr) sets the acceptance conditions: nested files must stay human-named weights (config list or explicit CLI name), never silent always-on **/AGENTS.md auto-discovery that widens writes without naming scope; evidence must be scoped per subtree and association via tool paths/cwd must stay deterministic, with anything best-effort opt-in (accuracy over coverage); keep two-session verbatim evidence, apply as the only writer, and per-file budget; today's "consolidate separate files" warning is the wrong diagnosis for intentional nested weights, so a fix should name the nested surface clearly; the contract class is opt-in if unconfigured runs keep today's single-primary behavior exactly; it is not a restore (single-primary plus separate warning stays intentional for the AGENTS/CLAUDE pair).

What Changed

  • Add opt-in nestedMemoryFiles with deterministic, subtree-scoped session attribution and routing of new instructions to the most specific named file.
  • Analyze and synthesize nested files separately, with per-file budgets, while combining their edits into one proposal and apply review.
  • Show nested-file status and budgets in the CLI and apply UI; preserve single-primary behavior when no nested files are configured.

Risk Assessment

🚨 High: The nested evidence-selection rule conflicts with the stated touched-subtree criterion, and two routing paths can silently select the wrong training surface.

Testing

Targeted tests passed. Live CLI runs identified both sessions, attributed one to the named API subtree, displayed its separate budget, preserved root-only behavior when unconfigured, and refused invalid targets. The short live sessions were skipped before model analysis, so no proposal or apply surface was produced; there is consequently no rendered review artifact.

  • Live validation: ⚠️ inconclusive - 3 of 5 scenarios driven live against the product
Scenario Result Live Evidence
Name an API memory file and see its own budget and API-only session attribution ✅ pass live live-analyze.txt; live-boundaries.txt
Leave nested memory unconfigured and retain the root-only surface ✅ pass live live-boundaries.txt
Name a missing nested file or request an unauthorized target without widening training ✅ pass live live-boundaries.txt; live-target.txt
Train scoped and cross-cutting lessons, review one per-file-budget proposal, and apply accepted edits ⏸️ untested no The isolated live sessions were too short for model analysis, and this phase had no authenticated model harness for a longer analysis and synthesis run. Provide authenticated harness access and suffic…
Have synthesis refuse a direct write to another named memory file ⏸️ untested no No authenticated, controllable live synthesis harness was available to attempt the direct write. Provide one in the isolated environment and rerun.
Evidence: Live subtree attribution
· analyzing 2 transcript(s) with pi effort=medium at jobs=4
  2/2 analyzed
· apps/api/AGENTS.md is a nested memory file: 1 of 2 session(s) worked under apps/api/
· analyzing 1 transcript(s) with pi effort=medium at jobs=4
  1/1 analyzed

analyzed against AGENTS.md (0 instructions, 4 tok)
  0 newly analyzed · 0 cached · 2 skipped (too short) · 0 failed
analyzed against apps/api/AGENTS.md (nested, 0 instructions, 4 tok) from 1 session(s) under apps/api/
  0 newly analyzed · 0 cached · 1 skipped (too short) · 0 failed
Evidence: Live configuration and missing-file boundaries
=== configured nested status ===
repo ~/.no-mistakes/worktrees/6e33ff56f8c4/01M3H5CG3YM6EMDPDV2GNB5VWB/.local-nested-live/repo

BUDGET (always-loaded)
  AGENTS.md          [................................] 4 / 5,000 tok · 0 instructions
  apps/api/AGENTS.md [................................] 4 / 1,000 tok · 0 instructions · nested, loads under apps/api/

CACHE
  scan cache      2 file(s) fingerprinted
  evidence        0 ok · 2 skipped · 0 failed
  rejections      0 remembered

PROPOSAL
  none yet - run `backpass`

MODELS
  analysis        auto - 7 candidates, none probed yet
  synthesis       auto - 7 candidates, none probed yet
=== unconfigured status ===
repo ~/.no-mistakes/worktrees/6e33ff56f8c4/01M3H5CG3YM6EMDPDV2GNB5VWB/.local-nested-live/repo

BUDGET (always-loaded)
  AGENTS.md      [................................] 4 / 5,000 tok · 0 instructions

CACHE
  scan cache      2 file(s) fingerprinted
  evidence        0 ok · 2 skipped · 0 failed
  rejections      0 remembered

PROPOSAL
  none yet - run `backpass`

MODELS
  analysis        auto - 7 candidates, none probed yet
  synthesis       auto - 7 candidates, none probed yet
=== unconfigured root scan ===
repo · 1 worktree(s) · since all

HOST   HARNESS  SCANNED  MATCHED  SELF  CACHED  NOTE
local  pi       2        2        0     2

2 transcript(s) associated with this repo · tier1 2 (exact) · tier1.5 0 (sibling clone) · tier2 0 (remote) · tier3 0 (best-effort) · interactive 0 · non-interactive 2

HOST   HARNESS  SESSION  KIND             WHEN   SIZE  TIER  HOW
local  pi       web-one  non-interactive  today  1KB   t1    cwd is worktree ~/.no-mistakes/worktrees/6e33ff56f8c4/01M3H5CG3YM6EMDPDV2GNB5VWB/.local-nested-live/repo
local  pi       api-one  non-interactive  today  1KB   t1    cwd is worktree ~/.no-mistakes/worktrees/6e33ff56f8c4/01M3H5CG3YM6EMDPDV2GNB5VWB/.local-nested-live/repo
=== named missing nested file ===
repo ~/.no-mistakes/worktrees/6e33ff56f8c4/01M3H5CG3YM6EMDPDV2GNB5VWB/.local-nested-live/repo

BUDGET (always-loaded)
  AGENTS.md              [................................] 4 / 5,000 tok · 0 instructions
  apps/missing/AGENTS.md nested - does not exist, not trained

CACHE
  scan cache      2 file(s) fingerprinted
  evidence        0 ok · 2 skipped · 0 failed
  rejections      0 remembered

PROPOSAL
  none yet - run `backpass`

MODELS
  analysis        auto - 7 candidates, none probed yet
  synthesis       auto - 7 candidates, none probed yet
=== restore ===
Evidence: Live target refusals
=== unnamed nested target refused ===
error --target "apps/web/AGENTS.md" is not a configured memory file or a loaded skill in this repo
  memory files: AGENTS.md, CLAUDE.md · skills: (none)
exit=1
=== named nested target ===
error --target apps/api/AGENTS.md is a nested memory file, which a run over the whole surface trains
  run backpass without --target: each nested memory file learns from the sessions that worked under its directory
exit=1
- Outcome: ⚠️ 1 warning across 3 runs (11m42s)

Known limitations / follow-ups

  • Shared cross-pass gap identity between root and nested ledgers. Each trained file is analyzed in its own pass with its own gap ledger. Exact session-and-quote matching keeps the common case of a cross-cutting lesson out of nested passes, but two distinct lessons citing the same quote can be conflated, and a root-owned lesson can still be proposed for both the root and a nested file when the passes quote it differently. A gap identity shared across ledgers is the follow-up.
  • Per-hunk evidence for edits that add instructions for several scopes. Routing judges an edit's additions by the combined evidence behind the edit, so a single edit adding one API-only and one web-only instruction is treated as cross-cutting. Per-hunk evidence, or requiring independently evidenced addition edits, is the follow-up.
  • Strict all-in-subtree attribution. A session trains a nested file only when all of its in-repo work paths are inside that file's directory; a session that edits apps/api/src/orders.ts and also reads README.md feeds only the root file. This is a deliberate accuracy-over-coverage choice and is open to maintainer preference.
  • Out of scope here: --target on a nested file is refused rather than supported; sessions collected over ssh are not attributed to a subtree; nested files are not offered skill extraction; the live TUI does not label nested passes separately; a failed nested synthesis fails the whole run.

apply remains the human gate for all of the above: every proposed edit is labeled with its target file, and nothing is written without review.

Closes #158

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

⚠️ **Review** - 3 issues (1 error, 2 warnings)
  • 🚨 src/nested.js:330 - The required criterion says to feed a nested file evidence from sessions that “touched its subtree.” Here, paths.every(...) excludes a session that edits apps/api/src/orders.ts and also reads README.md. src/commands/analyze.js:132 therefore omits it from API analysis, and src/nested.js:349 also prevents its evidence from owning an API lesson. This is the diff’s stricter, all-work-must-stay-under-the-directory rule; confirm whether that narrower policy is intended or scope evidence from sessions that touch the subtree.
  • ⚠️ src/proposal.js:591 - A root edit with one rewrite hunk and a separate pure-addition hunk backed by two API-only sessions has onlyAdds === false, so the addition bypasses routing and can land in the always-loaded root. The same bypass applies to an ancestor nested file when a deeper file owns the addition. Check addition hunks within mixed edits and require the misplaced addition to be split into its owning file.
  • ⚠️ src/nested.js:274 - For an associated session without a recorded cwd, a tool call with relative workdir: "apps/api" and relative path: "src/orders.ts" resolves against the backpass process cwd, not a known session checkout. It can silently attribute work to the current repo’s API subtree despite an unknown original base. Treat relative tool paths and workdirs as unplaceable unless the session supplies an absolute base.

🔧 Fix applied.
7 issues (2 errors, 5 warnings) still open:

  • 🚨 src/nested.js:330 - The required criterion says to feed a nested file evidence from sessions that “touched its subtree.” Here, paths.every(...) excludes a session that edits apps/api/src/orders.ts and also reads README.md. src/commands/analyze.js:132 therefore omits it from API analysis, and src/nested.js:349 also prevents its evidence from owning an API lesson. This is the diff’s stricter, all-work-must-stay-under-the-directory rule; confirm whether that narrower policy is intended or scope evidence from sessions that touch the subtree.
  • ⚠️ src/proposal.js:591 - A root edit with one rewrite hunk and a separate pure-addition hunk backed by two API-only sessions has onlyAdds === false, so the addition bypasses routing and can land in the always-loaded root. The same bypass applies to an ancestor nested file when a deeper file owns the addition. Check addition hunks within mixed edits and require the misplaced addition to be split into its owning file.
  • ⚠️ src/nested.js:274 - For an associated session without a recorded cwd, a tool call with relative workdir: "apps/api" and relative path: "src/orders.ts" resolves against the backpass process cwd, not a known session checkout. It can silently attribute work to the current repo’s API subtree despite an unknown original base. Treat relative tool paths and workdirs as unplaceable unless the session supplies an absolute base.
  • 🚨 src/synthesize.js:629 - With nested files configured, synthesis fingerprints only the file it is editing and staged skills. A nested synthesis agent given write access can write root AGENTS.md directly without assertRepoUntouched detecting it; the later root pass can then propose against that unreviewed change. The root pass likewise does not detect a direct write to a nested file. This leaves the required apply-only writer invariant reachable through both passes (src/commands/propose.js:147,151). Fingerprint every named memory file at the shared synthesis boundary.
  • ⚠️ src/nested.js:335 - Two distinct gaps may cite the same verbatim quote in the same API sessions. If one also has a web sighting, rootOwnedGaps records it as root-owned; the API-only gap then matches its API session-and-quote subset and is discarded from the nested fold (src/fold.js:272). The proposal gate makes the same mistaken root assignment (src/proposal.js:601). Include gap identity in the ownership comparison so one lesson cannot suppress another.
  • ⚠️ src/proposal.js:591 - Round 1's fix checks whether any hunk is a pure addition, but still assigns all addition hunks the owner of the edit's combined evidence (src/proposal.js:592-603). A root edit adding separate API and web instructions, supported respectively by API and web sessions, has a combined root owner and passes despite both additions belonging in nested files. The smallest precise remedy needs hunk-specific evidence or an authorized restriction requiring independently evidenced addition edits; that change to the annotation contract needs user approval.
  • ⚠️ src/nested.js:112 - The added rootNames expansion treats every root memoryFiles basename as a nested sibling memory file. With root memoryFiles: ["README.md"] and a named apps/api/AGENTS.md, an ordinary apps/api/README.md is reported as a separate file to consolidate. No stated requirement needs this extra matching path: the required per-directory AGENTS.md/CLAUDE.md pointer check can use that pair. Remove the root-basename expansion.

🔧 Fix applied.
6 issues (1 error, 5 warnings) still open:

  • 🚨 src/nested.js:330 - The required criterion says to feed a nested file evidence from sessions that “touched its subtree.” Here, paths.every(...) excludes a session that edits apps/api/src/orders.ts and also reads README.md. src/commands/analyze.js:132 therefore omits it from API analysis, and src/nested.js:349 also prevents its evidence from owning an API lesson. This is the diff’s stricter, all-work-must-stay-under-the-directory rule; confirm whether that narrower policy is intended or scope evidence from sessions that touch the subtree.
  • ⚠️ src/nested.js:274 - For an associated session without a recorded cwd, a tool call with relative workdir: "apps/api" and relative path: "src/orders.ts" resolves against the backpass process cwd, not a known session checkout. It can silently attribute work to the current repo’s API subtree despite an unknown original base. Treat relative tool paths and workdirs as unplaceable unless the session supplies an absolute base.
  • ⚠️ src/nested.js:335 - Two distinct gaps may cite the same verbatim quote in the same API sessions. If one also has a web sighting, rootOwnedGaps records it as root-owned; the API-only gap then matches its API session-and-quote subset and is discarded from the nested fold (src/fold.js:272). The proposal gate makes the same mistaken root assignment (src/proposal.js:601). Include gap identity in the ownership comparison so one lesson cannot suppress another.
  • ⚠️ src/proposal.js:591 - Round 1's fix checks whether any hunk is a pure addition, but still assigns all addition hunks the owner of the edit's combined evidence (src/proposal.js:592-603). A root edit adding separate API and web instructions, supported respectively by API and web sessions, has a combined root owner and passes despite both additions belonging in nested files. The smallest precise remedy needs hunk-specific evidence or an authorized restriction requiring independently evidenced addition edits; that change to the annotation contract needs user approval.
  • ⚠️ src/nested.js:112 - The added rootNames expansion treats every root memoryFiles basename as a nested sibling memory file. With root memoryFiles: ["README.md"] and a named apps/api/AGENTS.md, an ordinary apps/api/README.md is reported as a separate file to consolidate. No stated requirement needs this extra matching path: the required per-directory AGENTS.md/CLAUDE.md pointer check can use that pair. Remove the root-basename expansion.
  • ⚠️ src/nested.js:158 - With both apps/api/AGENTS.md and apps/api/db/AGENTS.md named, a DB session loads the API file too, but the DB evidence hash includes only root and DB. Changing the API file therefore leaves DB analysis cached against stale instructions. The sibling path at src/nested.js:191-197 and its caller at src/commands/analyze.js:155 also show the analyzer only the root file, so it can propose a gap already covered by API memory. Include every named ancestor that the session loads in both the hash and the also-loaded context.

🔧 Fix applied.
3 issues (1 error, 2 warnings) still open:

  • 🚨 src/nested.js:330 - The required criterion says to feed a nested file evidence from sessions that “touched its subtree.” Here, paths.every(...) excludes a session that edits apps/api/src/orders.ts and also reads README.md. src/commands/analyze.js:132 therefore omits it from API analysis, and src/nested.js:349 also prevents its evidence from owning an API lesson. This is the diff’s stricter, all-work-must-stay-under-the-directory rule; confirm whether that narrower policy is intended or scope evidence from sessions that touch the subtree.
  • ⚠️ src/nested.js:335 - Two distinct gaps may cite the same verbatim quote in the same API sessions. If one also has a web sighting, rootOwnedGaps records it as root-owned; the API-only gap then matches its API session-and-quote subset and is discarded from the nested fold (src/fold.js:272). The proposal gate makes the same mistaken root assignment (src/proposal.js:601). Include gap identity in the ownership comparison so one lesson cannot suppress another.
  • ⚠️ src/proposal.js:591 - Round 1's fix checks whether any hunk is a pure addition, but still assigns all addition hunks the owner of the edit's combined evidence (src/proposal.js:592-603). A root edit adding separate API and web instructions, supported respectively by API and web sessions, has a combined root owner and passes despite both additions belonging in nested files. The smallest precise remedy needs hunk-specific evidence or an authorized restriction requiring independently evidenced addition edits; that change to the annotation contract needs user approval.
⚠️ **Test** - 1 warning
  • ⚠️ Live proposal and apply validation could not proceed: the isolated Pi harness reported that it was not logged in. Provide model credentials to the isolated harness and rerun this phase to verify evidence routing and writes without a stand-in agent.
  • ⚠️ live validation verdict: inconclusive (4 of 6 scenarios were driven live against the product); untested: Train API and root from scoped two-session evidence, review one per-file-budget proposal, and apply both edits, Attempt a synthesis write to another named memory file and have the run refuse it
  • Live validation: ⚠️ inconclusive - 4 of 6 scenarios driven live against the product
Scenario Result Live Evidence
Name an API memory file and see it listed as a nested budget surface ✅ pass live Named nested status
Leave the nested list unset and see the API file remain separate, not silently trained ✅ pass live Unconfigured status
Scan isolated API and web sessions and see both associated with the monorepo ✅ pass live Strict session scan
Try to target a named nested file directly and receive a scope refusal ✅ pass live Nested target refusal
Train API and root from scoped two-session evidence, review one per-file-budget proposal, and apply both edits ⏸️ untested no The isolated Pi harness lacked model authentication. Provide credentials to that harness and rerun the live analysis, proposal, and apply flow.
Attempt a synthesis write to another named memory file and have the run refuse it ⏸️ untested no The isolated Pi harness lacked model authentication. Provide credentials to drive a real synthesis pass and attempt the cross-file write.
  • node --test test/nested-cli.test.js test/nested.test.js test/synthesize.test.js
  • backpass status with and without a named nested file in an isolated monorepo
  • backpass scan --harness pi --since all --host none --strict against isolated sessions
  • backpass analyze --target apps/api/AGENTS.md and backpass analyze --analysis-agent pi --analysis-model gpt-5.6-luna

🔧 Fix applied.
1 warning still open:

  • ⚠️ live validation verdict: inconclusive (5 of 7 scenarios were driven live against the product); untested: Train scoped and cross-cutting lessons, review one per-file-budget proposal, and apply accepted edits, Attempt a synthesis write directly to another named memory file and observe refusal
  • Live validation: ⚠️ inconclusive - 5 of 7 scenarios driven live against the product
Scenario Result Live Evidence
Leave nested files unconfigured and see only the root memory budget ✅ pass live default-status.json
Name nested files and see separate per-file budgets without treating API documentation as memory ✅ pass live nested-status.json; live-status.log
Leave a nested file in the legacy memoryFiles list or name a missing nested file and see the distinct diagnoses ✅ pass live legacy-status.json; missing-status.json
Target a nested file directly and receive a refusal rather than silently widening training ✅ pass live target-analyze.log
Scan an isolated monorepo corpus and associate its API and web sessions ✅ pass live live-scan.log
Train scoped and cross-cutting lessons, review one per-file-budget proposal, and apply accepted edits ⏸️ untested no The isolated Pi harness required authentication before analysis. Provide authorized model credentials to that harness and rerun live validation.
Attempt a synthesis write directly to another named memory file and observe refusal ⏸️ untested no Live synthesis requires the same unavailable Pi harness authentication; provide authorized model credentials and rerun.
  • node bin/backpass.js status --json with no nested configuration, then with named nested files and a missing file
  • node bin/backpass.js analyze --target apps/api/AGENTS.md --harness pi --since all
  • node bin/backpass.js scan --harness pi --since all against four isolated sessions
  • node bin/backpass.js analyze --harness pi --since all --analysis-agent pi --jobs 1
  • node --test test/nested-cli.test.js test/nested.test.js test/synthesize.test.js

🔧 Fix applied.
1 warning still open:

  • ⚠️ live validation verdict: inconclusive (3 of 5 scenarios were driven live against the product); untested: Train scoped and cross-cutting lessons, review one per-file-budget proposal, and apply accepted edits, Have synthesis refuse a direct write to another named memory file
  • Live validation: ⚠️ inconclusive - 3 of 5 scenarios driven live against the product
Scenario Result Live Evidence
Name an API memory file and see its own budget and API-only session attribution ✅ pass live live-analyze.txt; live-boundaries.txt
Leave nested memory unconfigured and retain the root-only surface ✅ pass live live-boundaries.txt
Name a missing nested file or request an unauthorized target without widening training ✅ pass live live-boundaries.txt; live-target.txt
Train scoped and cross-cutting lessons, review one per-file-budget proposal, and apply accepted edits ⏸️ untested no The isolated live sessions were too short for model analysis, and this phase had no authenticated model harness for a longer analysis and synthesis run. Provide authenticated harness access and suffic…
Have synthesis refuse a direct write to another named memory file ⏸️ untested no No authenticated, controllable live synthesis harness was available to attempt the direct write. Provide one in the isolated environment and rerun.
  • node --test test/nested-cli.test.js test/nested.test.js test/synthesize.test.js
  • backpass scan --harness pi --since all --host none in an isolated monorepo with API and web sessions
  • backpass analyze --harness pi --since all --host none --analysis-agent pi --limit 2
  • backpass status --host none with configured, unconfigured, and missing nested files
  • backpass analyze --target apps/web/AGENTS.md and --target apps/api/AGENTS.md
✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.

The signal round-trip test spawns two node processes and forwards a
signal between them. Alone it finishes well under a second, but with
the whole suite running in parallel it intermittently exceeded the
2-second bound. A wrapper that swallows the signal never exits, so a
10-second bound still fails it.
Monorepos keep per-directory memory files (apps/api/AGENTS.md) that
harnesses load on top of the root file only when a session works under
that directory. backpass only trained the root file, and a nested file
listed in memoryFiles was misreported as a separate file to consolidate.

nestedMemoryFiles (project scope, default []) names each nested file a
human wants trained; nothing is discovered. Each named file is a weight
of its own:

- evidence: only sessions that worked under its directory, judged
  against it with the root file shown as already loaded, in its own
  state under .backpass/nested/
- attribution: deterministic, from the session cwd and the paths tool
  calls name in structured fields (plus apply_patch file headers),
  resolved against known checkouts of the repo; ssh sessions and paths
  outside every checkout stay unattributed and feed only the root file
- routing: a new instruction lands in the most specific named file
  whose directory every session behind it worked in; cross-cutting or
  unattributed evidence stays with the root file, enforced by fold and
  by a buildProposal gate on pure additions
- budget: nestedBudgetTokens (budgetTokens when unset), re-checked at
  apply against the accepted subset
- review: one proposal and one apply surface, each nested edit labeled
  with its file; the 2-session floor, verbatim evidence, and apply as
  the only writer are unchanged

The per-directory pointer model matches the root: a sibling CLAUDE.md
that imports the nested AGENTS.md is fine, and a nested entry that is
only a pointer is refused by name. --target on a nested file is refused
and --memory-file runs never train nested files. With nestedMemoryFiles
unset, runs are unchanged.

Closes kunchenguid#158
@greptile-apps

greptile-apps Bot commented Sep 27, 2026 •

Copy link
Copy Markdown

RetriggerConfidence Score: 2/5

[High risk] Adds nested memory file training to monorepo support.

The PR is not safe to merge until nested evidence selection and instruction ownership preserve the intended per-directory boundaries.

Reviews (1) · Last reviewed commit: "no-mistakes(document): Clarify nested me..."

Comment thread src/nested.js
Comment on lines +338 to +339
export function workedUnder(paths, dir) {
return Array.isArray(paths) && paths.length > 0 && paths.every((relative) => isWithin(relative, dir));

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 API sessions excluded A session that edits apps/api/src/orders.ts but also reads the root README.md fails this every-path check, so it is left out of API analysis. The accepted intent is to train a nested file from sessions that touched its subtree; this loses API evidence and can leave an API-specific lesson in the always-loaded root.

Comment thread src/proposal.js
Comment on lines +591 to +603
if (routing && !preservesAlwaysLoaded(edit.kind) && hunks.some((h) => h.added > 0 && h.removed === 0)) {
const sessions = [
...new Set(
edit.evidence.map((item) => summary?.sourceSessions?.[normalizeSourceLabel(item.source)]).filter(Boolean),
),
];
const sightings = edit.evidence.map((item) => ({
sessionId: summary?.sourceSessions?.[normalizeSourceLabel(item.source)],
quote: item.text,
}));
const owner = rootOwnsGap(sightings, routing.rootOwnedGaps)
? routing.rootPath
: (routing.ownerOf(sessions) ?? routing.rootPath);

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Separate additions reach root When one root edit contains an API-only addition and a separate web-only addition, this check combines their evidence before deciding where the edit belongs. The combined evidence appears cross-cutting, so both additions pass into the always-loaded root even though each belongs in its named nested file.

Knowledge Base Used: Proposal synthesis

Comment thread src/nested.js
Comment on lines +342 to +350
export function rootOwnsGap(items, rootOwnedGaps = []) {
return (
items.length > 0 &&
rootOwnedGaps.some((sightings) =>
items.every((item) =>
sightings.some((sighting) => sighting.sessionId === item.sessionId && sighting.quote === item.quote),
),
)
);

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Distinct gaps get conflated Two different mistakes can cite the same quote in the same API sessions. If one mistake also occurs outside API, its root-owned quotes match the API-only gap here because the gap itself is not compared. The nested fold then drops the separate API lesson, and the proposal gate can reject adding it to the API file.

Knowledge Base Used: Proposal synthesis

@kunchenguid
kunchenguid merged commit 87a79fb into kunchenguid:main Sep 27, 2026
5 of 6 checks passed
@kunchenguid

Copy link
Copy Markdown
Owner

Speaking as Kun's firstmate: this is merged. Thank you @Athleticsmind — really appreciate you taking the time on this.

contract-class: opt-in — nestedMemoryFiles defaults to []; unconfigured runs keep today's single-primary behavior exactly. Not restore (single-primary + separate warning stays intentional for the AGENTS/CLAUDE pair). Not new-default (no silent **/AGENTS.md discovery).

VISION per-rule

  • Purpose / owns the backward pass — aligns (nested project weights still the backward pass into human-named surfaces).
  • Evidence is the only currency — aligns (subtree attribution; ssh/unplaceable → root only; accuracy over coverage).
  • The human owns the weights — aligns (config list only; no silent widen; apply remains the only writer; per-edit review labeled by target file).
  • Nothing the model says is taken on faith — aligns (measurement/apply gate unchanged; named memory files fingerprinted across nested synthesis).
  • The budget is the constraint — aligns (per-file nested budget re-checked at apply).
  • It reads what you already have and owns nothing — aligns (deterministic tool-path/cwd attribution; best-effort left out).
  • Failure is loud and named — aligns (missing nested named; nested synthesis failure named by file).
  • Scope — aligns (project-scope only; no user-level write from a project run).
  • Closing resist/align — aligns (named + scoped; no write-widen).

Security: clean (paths confined to known checkouts; no shell-text attribution; apply gate intact).
CI: fork workflows approved; CI/Guard/no-mistakes green. Greptile fail ignored (not used for this repo). Strict all-in-subtree attribution and mixed-edit/gap-identity follow-ups are documented known limitations, not blockers.

Closes #158.

MrGTV-love added a commit to MrGTV-love/backpass that referenced this pull request Oct 20, 2026
* fix(acpx): report acpx budget kills as timeouts, not empty-output (kunchenguid#145)

Closes kunchenguid#143.

Names acpx --timeout blank-quiet kills as timedOut on execOneShot and sessionPrompt (0.9 budget floor); leaves genuine fast blanks as empty-output; leaves synthesis openSession/prompt() alone.

* chore(main): release backpass 0.1.27 (kunchenguid#149)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* fix(redact): stop secret redaction from eating benign tool arguments (kunchenguid#153)

* fix(redact): anchor ASSIGNMENT identifier, bound value, skip numerics

Benign telemetry names that merely contain a secret word
(max_output_tokens, prompt_tokens, token_count) no longer match:
the identifier must now end with the secret word. The value stops
at whitespace, a quote, or a comma, so comma-joined JS-literal
argument lists survive, and purely numeric values are left alone.
Genuine assignments (API_KEY=, GITHUB_TOKEN:, password=) still redact.

* no-mistakes(review): restore secret-word breadth, drop numeric redaction bypass

* no-mistakes(review): redact unterminated-quote secrets, pin value-bound tests

* no-mistakes(document): format new redaction tests, docs already accurate

* no-mistakes(document): align redact comment with final value pattern

* fix(redact): keep quoted values on one line

An unterminated quote no longer redacts across newlines: the
closed-quote alternatives exclude \n, so a truncated assignment
falls through to the bounded unquoted form and following trace
text survives.

* no-mistakes(ci): ci-4 fixed in src/redact.js. Invariant: a fully closed quoted secret value must be redacted whole, even across newlines, without a stray later quote swallowing benign prose. Change: the ASSIGNMENT pattern now tries two newline-spanning quoted alternatives first — `"([^"]{8,})"` and `'([^']{8,})'`, each gated by a lookahead `(?=[ \t]*(?:$|[,;})\]]))` with the `m` flag — before the existing single-line quoted alternatives and the bare fallback. Both quote directions (double and single) got the same edit, and the replace callback's group chain became `dqMulti ?? sqMulti ?? dq ?? sq ?? bare` so every alternative still feeds the already-redacted guard. Traced: `PASSWORD="abcdefgh\nijklmnop"` (closing quote at end of line) now masks both lines. `DB_PASSWORD="hunter2hunter2secret\nRun tests and note "done"` still fails the lookahead (the only candidate closing quote is followed by `done"`), so it falls through to the bare form and masks only the first-line value, leaving the Run-tests line visible. The single-line quoted alternatives remain reachable (e.g. `password="secretvalue" trailing words`). Comma/unterminated-quote cases from earlier rounds unchanged. Test added in test/distill.test.js next to the existing unterminated-quote case. `pnpm run check` exits 0: 692 tests pass, lint, format:check and typecheck clean

---------

Co-authored-by: firstmate <firstmate@local>

* docs: trim always-loaded agent guidance (kunchenguid#156)

* docs(agents): trim AGENTS.md to every-session invariants with pointers

AGENTS.md is loaded into every agent session. Keep only hard invariants and
orientation as short lines pointing at the source file that owns the detail.

Before: 406 lines, 4663 words, 33498 bytes. After: 87 lines, 907 words, 7001 bytes.
No detail moved to docs: every removed explanation already lives in the named
module's header/inline comments and tests. CLAUDE.md stays the @AGENTS.md pointer.

* no-mistakes(ci): Corrected AGENTS.md to state that skillSearchPaths skills are withheld from staging but remain fingerprinted, and restored the >= 20% placement table as prompt guidance rather than a buildProposal gate. pnpm run check passed (692 tests)

* chore(main): release backpass 0.1.28 (kunchenguid#155)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* feat: train named nested memory files in monorepos (kunchenguid#159)

* test(harness-invoke): widen the signal-forwarding budget

The signal round-trip test spawns two node processes and forwards a
signal between them. Alone it finishes well under a second, but with
the whole suite running in parallel it intermittently exceeded the
2-second bound. A wrapper that swallows the signal never exits, so a
10-second bound still fails it.

* feat(memory): train nested memory files in a monorepo

Monorepos keep per-directory memory files (apps/api/AGENTS.md) that
harnesses load on top of the root file only when a session works under
that directory. backpass only trained the root file, and a nested file
listed in memoryFiles was misreported as a separate file to consolidate.

nestedMemoryFiles (project scope, default []) names each nested file a
human wants trained; nothing is discovered. Each named file is a weight
of its own:

- evidence: only sessions that worked under its directory, judged
  against it with the root file shown as already loaded, in its own
  state under .backpass/nested/
- attribution: deterministic, from the session cwd and the paths tool
  calls name in structured fields (plus apply_patch file headers),
  resolved against known checkouts of the repo; ssh sessions and paths
  outside every checkout stay unattributed and feed only the root file
- routing: a new instruction lands in the most specific named file
  whose directory every session behind it worked in; cross-cutting or
  unattributed evidence stays with the root file, enforced by fold and
  by a buildProposal gate on pure additions
- budget: nestedBudgetTokens (budgetTokens when unset), re-checked at
  apply against the accepted subset
- review: one proposal and one apply surface, each nested edit labeled
  with its file; the 2-session floor, verbatim evidence, and apply as
  the only writer are unchanged

The per-directory pointer model matches the root: a sibling CLAUDE.md
that imports the nested AGENTS.md is fine, and a nested entry that is
only a pointer is refused by name. --target on a nested file is refused
and --memory-file runs never train nested files. With nestedMemoryFiles
unset, runs are unchanged.

Closes kunchenguid#158

* no-mistakes(review): Fix nested evidence routing and per-directory memory checks

* no-mistakes(review): Correct nested attribution for tool paths and cwd

* no-mistakes(review): Guard nested apply scope and independent budgets

* no-mistakes(review): Allow nested proposals past unchanged over-budget roots

* no-mistakes(review): Share root-owned gap decisions with nested passes

* no-mistakes(test): Document root-owned gap route type in foldEvidence

* no-mistakes(document): Clarify nested targeting and apply scope in README

* no-mistakes(review): Require nested synthesis before exempting root budget

* no-mistakes(document): Clarify nested memory budgets and skipped training

* no-mistakes(review): Fix mixed-edit routing and relative path attribution

* no-mistakes(review): Protect named memory files during synthesis

* no-mistakes(review): Include nested ancestors in analysis context and hash

* no-mistakes(test): Fix tuple typing and formatting; focused tests pass, checks prohibited

* no-mistakes(test): Adjust nested formatting; focused tests pass, formatter verification prohibited

* no-mistakes(document): Clarify nested memory documentation and pass lint checks

* doc: Revise explanation of AGENTS.md and backpass functionality

Updated README to clarify the role of AGENTS.md and backpass.

* doc: Update README to include skills in backpass description

* fix(discovery): include Hermes TUI sessions with absolute cwd (kunchenguid#146)

* fix(hermes): discover TUI sessions with absolute cwd

* no-mistakes(document): Confirm Hermes documentation and format changed test

* feat(apply): keep reworded rejected edits suppressed and record optional reject reasons (kunchenguid#162)

* feat: remember rejected gaps and instruction units across rewordings

Rejecting an add, rewrite, or remove now stores measured gap and
instruction identities so a later proposal of the same decision in
different bytes stays suppressed until more sessions arrive. Extract
and move stay hunk-key only. The browser apply surface can attach an
optional reason; terminal review does not prompt for one.

* no-mistakes(review): Key rejections on unit hashes and absorbed gap aliases

* chore(main): release backpass 0.1.29 (kunchenguid#160)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* docs: trim trailing space in README (kunchenguid#170)

Prettier format:check fails on main because the scientific rigor
sentence has a trailing space after the period.

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Kun Chen <kunchenguid@users.noreply.github.com>

* test: isolate the suite's config home and temp fixtures per test process (kunchenguid#167)

* test: isolate the suite from the developer's personal config home

loadConfig layers $XDG_CONFIG_HOME/backpass/config.json under every run, and
the suite never set XDG_CONFIG_HOME, so a contributor with personal ladders in
~/.config/backpass/config.json saw 14 failures in agents.test.js and
config.test.js (0 with an empty config home). A setup module loaded with
--import gives every test process an empty config home of its own; tests
that need one still set and restore XDG_CONFIG_HOME themselves.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* no-mistakes(ci): I fixed the Greptile P2 in `test/helpers/hermetic-config.js`, following your instructions. Before this, the temporary config home was only removed on a normal `exit`, so a test run stopped with SIGINT or SIGTERM left it behind in the temp directory. **Change:** the existing `exit` cleanup is unchanged, now through a shared `removeConfigHome` function. I added one `process.once` listener each for SIGINT and SIGTERM. When a signal arrives, the listener: - deletes the directory (recursive, force, and any error is ignored); - removes its own listeners for both signals; - re-sends the same signal with `process.kill(process.pid, signal)`, so the process still ends the normal way for that signal. Every process that loads the helper through `--import` gets this, and nothing else changed. **Verification:** - **Signals:** I started a process with the helper loaded, checked its config directory existed, then sent it a signal. After SIGTERM it exited with 143 and the directory was gone; after SIGINT it exited with 130 and the directory was gone. - **Tests:** `npm test` passed 738 of 738, with 0 failed and 0 cancelled. - **Lint, types, formatting:** eslint and `tsc --noEmit` pass, and prettier passes on the changed file. - **`pnpm run check`:** this still fails at `format:check`, but only on `README.md`, which this change doesn't touch. Open PR kunchenguid#166 already fixes that file. I have not committed the change; it is left in the worktree for the pipeline

* test: give each test process its own temp root so fixtures stop leaking

Tests create fixtures with fs.mkdtempSync(os.tmpdir()) and never remove them:
a full run left about 1,275 backpass-* directories (~130 MiB) in the system
temp dir, over 10,000 after a day of runs. The hermetic helper now points
TMPDIR (read by os.tmpdir() on every call, inherited by spawned CLI children)
at one per-process root and removes it on exit, SIGINT and SIGTERM. The config
home moves inside that root. Full suite 743/743 and 0 directories left behind.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test: make the temp-root cleanup best-effort and survive read-only fixtures

Now that one root holds every fixture, a test that fails before restoring a
0o555 directory (apply-cli, proposal, user-apply tests) made rmSync throw
EACCES inside the exit listener: exit 1, an unrelated stack trace, and the
root left behind. Removal now restores write permission and retries, and never
throws from exit or signal handlers. Old vs new on that scenario: exit 1 +
stack + root left, vs exit 0 + clean + removed. Full suite 743/743, 0 leaked.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* fix(acpx): retry timed-out adapter verification instead of aborting the run (kunchenguid#168)

* fix(acpx): a timed-out adapter-config check fails the call, not the run

verifyHarnessInvocation read a timed-out `acpx config show` (code null) as a
permanent verification failure and raised a UserError, which analyze.js
rethrows: one slow node start on a loaded host (load average ~30) rejected the
analysis pool at 23/100 with advice to upgrade acpx. A timeout says nothing
about the adapter map, so it is now handled before generic exits: retried once
with three times the budget, then raised as a timed-out AcpxError, which the
analysis stage records per transcript and retries on the next run. The budget
is injectable (verifyTimeoutMs) so the tests do not wait on real timeouts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* no-mistakes(review): Let pinned agents' timed-out acpx errors propagate unchanged

* no-mistakes(review): Print a retry hint for timed-out harness calls

* no-mistakes(review): Give verify timeouts their own accurate CLI hint

* no-mistakes(test): Isolate agents tests from host XDG config ladders

* no-mistakes(document): Strip trailing README whitespace failing format check

* no-mistakes(ci): Fixed Greptile finding ci-1 in AgentResolver.withFallthrough (src/agents.js), as the user instructed. The invariant: a timed-out AcpxError always propagates unchanged. It is never classified, demoted, or retried, for pinned and auto-picked agents alike. This matches the AGENTS.md rule that a timeout is never cached as a negative probe. Before the fix, a timeout whose stderr contained 'AUTH_REQUIRED' was classified as 'unauthenticated'. For a pinned pick, demote() returned false without throwing, so the loop called the same harness again with no end. For an auto-picked pick, the agent was demoted and written to the cache as a negative probe. The fix is one shared check at the top of the catch: `if (isAcpxError && err.timedOut) throw err;`. After that, the pinned conversion is back to plain `if (isAcpxError && pick.pinned)`, with the old `!err.timedOut` guard removed. All non-timeout pinned conversions (classifiable verdicts and 'failed unexpectedly') behave as before. The doc comment now says this. Tests added in test/agents.test.js: 1. A pinned synthesis pick whose fn throws a timed-out AcpxError with stderr 'AUTH_REQUIRED' rejects with that same error, and fn is called exactly once. 2. An auto-picked analysis candidate given the same error rejects with it once and is not demoted: resolve() still returns pi. Verification: - With src/agents.js reverted to the old code, the agents test file never finished inside 60s. That is the endless loop the finding describes. - With the fix, agents + config tests pass 52/52 under an isolated XDG_CONFIG_HOME. Lint, format:check and typecheck all exit 0. - Full `pnpm run check`: 741/743 pass. The 2 config.test.js failures come from this host's user config (~/.config/backpass sets a gpt-6-luna ladder), not from this PR: they fail identically with both changed files stashed and pass with an isolated XDG_CONFIG_HOME. - In a re-run under the isolated config home, the only failing test was gap-domain-cli ('an orchestration gap is counted…'). It exited 143 (SIGTERM after 30s) at host load average about 30, and passes 3/3 when run alone. That is host load, not this change. Changes are left uncommitted in the worktree for the outer executor

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* ci: add restricted workflow_dispatch path to republish npm 0.1.29 (kunchenguid#172)

Allow a one-off trusted-publishing republish of the already-tagged
backpass@0.1.29 release without bumping the version or creating a new
release-please release.

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Kun Chen <kunchenguid@users.noreply.github.com>

* chore(main): release backpass 0.1.30 (kunchenguid#171)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* fix(agents): treat "does not support this model" as model-unavailable (kunchenguid#185)

* fix(agents): fall through when the harness is too old for the model

An adapter can advertise a model that its bundled harness cannot call.
The provider then rejects the first prompt with "does not support this
model". Classify it as model-unavailable so withFallthrough demotes the
candidate, caches the negative verdict, and tries the next rung.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(agents): label model-unavailable as "model not available"

The verdict now also covers a listed model that the harness is too old
to call, so "model not advertised" was wrong for that case.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: zxliu <zxliu@scientechresearch.io>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix: isolate SQLite session raw transcripts and reclaim expired files (kunchenguid#178)

* fix(analyze): give a SQLite-store session a raw transcript of its own

The distilled trace ends with the raw transcript the analysis agent may
open when a claim needs the full text. For a file-backed store that is
the session's own file, and a session fetched over ssh already reads
from a cached file of its own events. For a local SQLite store (opencode,
hermes, cursor IDE) it was the store's path: one database holding every
session of every repository.

On a real run with an effortful model the analysis agent followed that
pointer and wrote SQL against a 560 MB OpenCode store to find its
session, turning minute-long calls into ones that hit the timeout - and
the database it was handed let it read sessions this repository never
associated.

For the length of the call, a local SQLite session now gets a file of
its own normalized events, one per line, under `.backpass/raw/` (mode
0600), and that is what the trace footer names. The file is removed when
the call returns, whatever the outcome. The new CLI test follows the
footer from inside a fake agent: it finds only that session's events,
never the database or another session, and the file is gone afterwards.

* no-mistakes(review): Fix SQLite raw-file lifecycle and adapter documentation

* no-mistakes(review): Reclaim orphaned raw transcripts without deleting live analysis files

* no-mistakes(review): Reclaim reused-PID raw files while preserving active analysis files

* no-mistakes(document): Clarify raw transcript scope and documentation ownership

* no-mistakes(ci): Fixed ci-1 by removing unsafe PID-based cross-run pruning from all analysis paths. Per-call and catchable-exit cleanup remain intact; SIGKILL leftovers now require manual removal, documented in README.md. Reproduced the same-PID deletion through the CLI before fixing it. All 100 relevant tests pass, along with targeted ESLint, Prettier, TypeScript, and git diff checks

* fix(analyze): reclaim a SQLite raw file once its lease runs out

A raw file left by an uncatchable exit such as SIGKILL was never removed,
leaving an unredacted session on disk. While its call runs, analysis now
renews the file's modification time every minute, and every analysis run
removes raw files untouched for 15 minutes from the root and nested state
directories. Liveness comes from the lease, never from a PID, so a live
analysis in another PID namespace keeps its file.

* no-mistakes(review): Remove redundant worker-level raw transcript reclamation

* no-mistakes(review): Simplify raw-file identity and restore reclamation test coverage

* no-mistakes(ci): Fixed both findings. Raised raw-file reclamation to 24 hours while preserving one-minute renewal. Directory scan failures now emit one warning and skip the affected directory without aborting analysis or healthy sibling cleanup. Updated README and comments. Reproduced failures through the CLI and added regressions covering clock skew, expired leases, and directory failures across root/nested, empty/cached runs. All 47 analysis tests passed, along with targeted ESLint, Prettier, typecheck, and git diff checks

* feat: support explicit per-edit decisions in backpass apply (kunchenguid#182)

* feat(apply): take a decision vector with --decisions

A reviewer who decided somewhere else, or a script someone chose to
trust with their own repo, had no way to hand apply its decisions: the
review surface needs a browser and --no-ui needs a terminal. --decisions
takes the vector the surface sends, e1=accepted e2=rejected:too-narrow,
parsed strictly so a typo writes nothing, and leaves unnamed edits
undecided. It is never a default, and every apply gate still runs.

* no-mistakes(ci): Fixed ci-1 by retaining every --decisions occurrence and validating them together through the existing strict parser. Earlier typos now stop all writes, and duplicate edit decisions across flags are refused. Updated help and README while preserving existing error messages. Added CLI regressions for invalid repeated flags, duplicate decisions, unchanged proposal/rejection state on failure, and successful combined decisions with rejection reasons. Reproduced the reported failure before the fix; afterward, all 144 relevant CLI, review-surface, scope, and writer tests passed. git diff --check passed

* fix(apply): refuse a blank --decisions vector among several

Joining repeated --decisions values let a blank one vanish next to a
valid one, so `--decisions " " --decisions "e1=accepted"` wrote e1 while
a blank vector alone is refused. Each supplied vector must now name an
edit on its own before the vectors are checked together.

* no-mistakes(document): Clarify scripted apply decisions and review defaults

* fix(state): skip unverifiable POSIX mode check for Windows user-profile state dirs (kunchenguid#195)

* fix: skip the POSIX mode read-back of the user state dir on Windows

On Windows, chmod only toggles the read-only attribute and stat always
reports 0o666, so State.ensure() could never verify 0o700 and every
user-scope command failed with "could not secure state directory ...
as mode 700 (got 666)". There, the directory's privacy comes from the
NTFS ACL it inherits from the user profile (SYSTEM, Administrators, the
user), so the read-back is skipped on win32 and kept everywhere else.

test/state-mode.test.js simulates Windows' stat and pins both sides:
the directory is accepted on win32 and still refused on other
platforms; each test goes red under the matching mutation of the guard.

Verified on Windows 11 / Node 26.7: `backpass scan --scope user`,
`status`, `analyze`, and `propose` now run end to end.

Fixes kunchenguid#115

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* no-mistakes(review): {"summary": "Narrow win32 mode-check skip to in-profile state dirs, warn otherwise"}

* no-mistakes(document): docs: reflect Windows --scope user state-dir support in README

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* feat(discovery): support additional OpenCode stores and 2.x transcripts (kunchenguid#181)

* fix(opencode): drop backpass's own OpenCode sessions, and only those

acpx drives opencode, and opencode is on both default ladders, so every
analysis and synthesis call backpass makes through it is filed in the
opencode store under the repo's cwd - a tier-1 exact match on the next
run. Self-session detection only reads transcript files, and an
opencode store is one SQLite database, so three things went wrong:

- The file-head check read the database's first 256 KiB. Its pages hold
  every session's text, so a backpass prompt stored near the start of
  the file marked every session in the store as backpass's own, and a
  prompt stored further in was missed.
- With the head check unable to see them, backpass's own opencode
  sessions were analyzed as corpus - backpass auditing itself, which the
  self-session sentinel exists to prevent.
- Each agent probe (`acpx opencode sessions new`) leaves an empty
  session under the repo's cwd, which was listed and matched at tier 1.

The opencode listing query now reads each session's first user text
part, the part an acpx prompt lands in, and marks the row `self` when it
opens with the sentinel; discovery drops those rows and counts them in
SELF. A session with no user text at all recorded nothing and is not
listed. Discovery never reads a SQLite database as a transcript head.
The new cases in `test/self-session.test.js` put a genuine session, a
self session and a probe's empty session in one small opencode store;
they fail on all three counts before this change.

* no-mistakes(review): Keep OpenCode discovery fail-soft on malformed JSON

* no-mistakes(review): Exclude descendants of OpenCode self sessions locally and remotely

* no-mistakes(test): Preserve attachment-only OpenCode sessions while excluding empty probes

* no-mistakes(document): Clarify OpenCode session exclusions and discovery safety documentation

* no-mistakes(lint): Fix self-session assertion formatting

* no-mistakes(ci): Fixed both findings. OpenCode self detection now checks the first user message, preserving attachment-opening conversations and their descendants when later messages quote backpass prompts. Remote probes respect the adapter's explicit classification. Self ancestry is computed once per discovery instead of querying ancestors per child. Both regressions failed before the fixes and pass afterward. Verification passed: 103 relevant tests, targeted ESLint and Prettier checks, TypeScript checking, and git diff --check. The hosted Greptile check was not rerun in this phase

* feat(opencode): read OpenCode 2 stores (session_v2 and session_message)

OpenCode 2 no longer writes `session`, `message` and `part`. Its store
keeps each session in `session_v2` and the whole history as one typed
row per message in `session_message`. On a 2.x-only store the adapter's
listing query failed with `no such table: session`, so the harness was
skipped with a warning and none of those sessions could be analyzed.

The adapter now reads both layouts from the one database:

- Listing reads `session_v2`, and dates a session by its newest message:
  `session_v2.time_updated` does not move as messages arrive, so a
  session resumed today was otherwise outside every window.
- An in-place upgrade copies each 1.x session into `session_v2` under
  the same id and leaves the 1.x tables behind; a session in both is
  listed once and read from `session_v2`, where it continues. A store
  with neither table is reported as drifted instead of read as empty.
- Reading maps each message type deliberately. User and assistant turns
  keep their text, and tool calls stay in the order the model made
  them, with their text output or error message as the result. Text the
  harness injected - instruction files (the memory file itself),
  continue notices, system notices, and compaction summaries of turns
  still stored above them - never becomes a turn. A background subagent
  or shell reports back in an injected message; that completion becomes
  the result of the call that started the job. A shell command or skill
  the person ran from the prompt is kept as a call.
- A 2.x session follows the same rules as a 1.x one: it is backpass's
  own when its first user message opens with the self-session sentinel
  or it descends from such a session, and it is not listed when it has
  no messages at all.

Golden fixtures pin both layouts: `test/fixtures/opencode-v2-store.json`
and `opencode-v1-store.json` carry the DDL of real 2.0.18 and 1.18.32
stores and rows modeled field-for-field on real ones.

* no-mistakes(review): Simplify OpenCode recording detection and clean temporary test homes

* no-mistakes(ci): Fixed ci-1 in src/tui/render.js by replacing the unmeasured "1 sqlite query" claim with "sqlite store" at the shared OpenCode/Hermes rendering boundary. Reproduced the incorrect rendered output before fixing it. Updated rendering tests to cover both harnesses at three terminal widths while preserving session counts. All 23 rendering tests, targeted ESLint and Prettier checks, and git diff --check pass. Discovery behavior is unchanged

* feat(opencode): read more OpenCode stores named in personal config

A person's OpenCode sessions do not always live in the one store
backpass reads. The Windows OpenCode app, for one, keeps its store on
the Windows side, so a backpass running under WSL beside it never sees
the sessions that did the work in this distro's checkouts.

`discovery.opencodeStores` names more stores to read beside the default
one: an OpenCode data directory (holding `opencode.db`) or a database
file, such as a copy of the Windows app's store.

Why a config list, when Claude, Codex, Hermes and Pi stores are
relocated through the environment (`CLAUDE_CONFIG_DIR`, `CODEX_HOME`,
...): OpenCode's own relocation variables, `XDG_DATA_HOME` and
`OPENCODE_DB`, would reach the OpenCode that backpass runs for analysis
through acpx, which passes backpass's environment on. That OpenCode
must keep its own store and login, or its sessions would land in, and
its credentials be looked up in, the store being read. A list read only
by the adapter changes what backpass reads and nothing else, the way
`cloneRoots` and `skillSearchPaths` add places to look without changing
anything they name.

- Like `discovery.hosts`, the list is personal configuration: a
  checked-in `.backpassrc.json` that sets it is refused by name, since
  it would point every contributor's backpass at a transcript store. In
  user scope it defaults to the global file's top-level list.
- Entries are absolute (`~/` is expanded); a relative entry is refused,
  because a transcript store is not a repository path.
- Each session is read from the store that listed it. A store named
  twice (the default one again, or through a link) is read once, and a
  session id already listed from an earlier store is skipped, so a
  copied session never counts twice toward `minGapEvidence`.
- A configured store that is missing or unreadable is named in a
  warning and skipped, and the rest of the run continues. The default
  store's failure stays the harness's failure, as before.
- Remote hosts are unaffected: the probe reads each host's default
  store.

* no-mistakes(review): Restore test environment and clean up OpenCode store fixtures

* no-mistakes(review): Remove personal-only settings from project configuration example

* no-mistakes(document): Clarify OpenCode store discovery documentation

* no-mistakes(ci): Fixed all four Greptile findings: missing SQLite stores now raise named errors, notice-only 2.x sessions are excluded, attachment-only user turns survive reading and distillation, and bounded discovery filters old sessions in SQLite while preserving resumed sessions and self ancestry. Reproduced nine failing regression cases before fixing. All 113 targeted tests pass, including CLI and remote-probe coverage; ESLint, formatting, TypeScript, and git diff checks also pass

* fix: stop scheduling analysis after a fatal worker error (kunchenguid#180)

* fix(analyze): stop starting model calls once a run has failed

Per-transcript failures are recorded and the run goes on; an error that
fails the run - a pinned agent failing in a way nothing can classify,
say - propagated from one analysis worker and stopped only that one.
`Promise.all` reported the failure at once while the other workers kept
taking transcripts and making model calls in the background, and the
process only exited when the whole queue was done. On a real run with a
pinned OpenCode agent the error was printed within minutes and the
process kept analyzing for over an hour, starting a new max-effort call
every time one finished.

The pool now stops handing out transcripts after the first fatal error.
Calls already in flight finish and keep their evidence, and the first
error is rethrown once every worker has stopped, so nothing is still
running when it surfaces. The new CLI test runs two jobs over five
sessions with a pinned agent whose first call fails: before this change
all five calls ran, now only the two in flight do.

* no-mistakes(document): Document fatal analysis shutdown and evidence retention

* no-mistakes(ci): Addressed ci-1 in test/analyze-stop-on-fatal.test.js. The CLI regression now returns non-empty, trace-backed evidence and verifies after exit that the second session's persisted record has status "ok" and retains the complete evidence payload. Existing assertions still verify fatal error reporting and no queued model calls. No production changes were needed. Verification passed: 6 relevant tests, targeted ESLint and Prettier checks, and git diff --check

* chore(main): release backpass 0.1.31 (kunchenguid#186)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* fix: recognize acpx exit code 3 as a timeout (kunchenguid#177)

* fix(acpx): name acpx's timeout exit (3) as a timeout

acpx 0.19 ends a model turn at its `--timeout` budget itself and exits
with its TIMEOUT code, 3 (`EXIT_CODES.TIMEOUT`), with only a zero-token
accounting line on stderr. backpass named a timeout only from its own
outer kill (`result.timedOut`, budget + 30s), which never fires once
acpx has exited, so the turn fell through to generic exit handling:

  error pinned analysis agent opencode (github-copilot/claude-opus-5.5)
  failed unexpectedly (acpx opencode session prompt failed (exit 3):
  [acpx] tokens: input=0 output=0 total=0)

With a pinned agent that stops the whole run after one long prompt.
Named a timeout, it is what the design already says it is: that
transcript is recorded as failed and retried next run, and the ladder
never falls through because a prompt ran long.

`ACPX_EXIT_TIMEOUT` is checked wherever `result.timedOut` is: the exec
one-shot, the session prompt, session create (so an adapter acpx gave up
starting is not reported as lacking session support), and the probe.

* no-mistakes(review): Distinguish acpx-owned timeout diagnostics from backpass budgets

* no-mistakes(review): Preserve timeout context and filter acpx accounting noise

* no-mistakes(review): Classify setup timeouts and preserve acpx error diagnostics

* no-mistakes(document): Clarify acpx timeout handling and retry documentation

* fix: prevent false session matches from Windows paths on POSIX (kunchenguid#173)

* fix(discovery): never resolve a Windows path against the POSIX cwd

A harness on Windows records `C:\work\repo` or a UNC path such as
`\\server\share\repo` (Windows tools also write `//server/share/repo`).
On macOS or Linux none of these is absolute, so `path.resolve` turned
them into a location under the process cwd - inside the checkout
backpass was started from - and every such session became a tier-1
match for whichever repo ran the scan (kunchenguid#164). The same resolution
also made a dead `C:/work/<repo-name>` a best-effort match, keyed
user-scope projects by the directory backpass ran in, and counted
Windows tool paths as in-repo work for nested memory files.

`src/discovery/paths.js` decides once whether a recorded path names a
place on this machine. Association, user-scope project keys and nested
attribution now refuse a Windows path on a POSIX host instead of
resolving it; a recorded remote still associates the session at tier 2.
On Windows the recorded spelling is already local, so nothing changes
there. The nested attribution cache version is bumped so cached
placements recompute.

Fixes kunchenguid#164

* no-mistakes(document): Clarify foreign-path handling and correct UNC safety rationale

* chore(main): release backpass 0.1.32 (kunchenguid#198)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* feat(discovery): add opt-in OMP session discovery to the Pi adapter (kunchenguid#151)

* feat(discovery): discover omp session transcripts in the pi adapter

omp (Oh My Pi) writes the same version:3 JSONL session format as pi under
~/.omp/agent/sessions/, but prepends a fixed-width {type:"title"} record so
the {type:"session", cwd} entry is line 2, and nests subagent transcripts
one level deeper at <escaped-cwd>/<session-id>/<Name>.jsonl.

- Add ~/.omp/agent/sessions as a default nested store root (omp honors
  PI_CODING_AGENT_DIR, so an env redirect still dedupes against it).
- classify() scans the first head lines for the session entry instead of
  requiring it on line 1.
- enumerate() descends one extra level inside nested roots.
- read() accepts model_change.model as a fallback for modelId.

* fix(discovery): share OMP corroboration across subagents

* no-mistakes(review): Guard ledger id migration, narrow fold overlay, nest OMP subagents

* no-mistakes(review): Drop graft cache files, guard ledger recording aliases

* no-mistakes(review): Make shared OMP domain vote order-independent

* no-mistakes(review): Link OMP subagents by nested path, share ledger prior merge

* no-mistakes(document): Sync OMP interaction and scan-cache doc comments

* chore: ignore local Graft cache

* no-mistakes(document): Sync OMP identity docs and rewrap long comments

* feat(discovery): make OMP session collection opt-in

* no-mistakes(review): Fix OMP observer deduplication, native ownership, and descendant classification

* no-mistakes(review): Correct observer canonicalization, native attribution, and selected provenance migration

* no-mistakes(review): Fix conflicting cross-host descendant family corroboration

* no-mistakes(review): Fix copied-descendant corroboration and historical ledger provenance

* no-mistakes(review): Reconcile copied families and preserve validated evidence provenance

* no-mistakes(document): Correct observer documentation and format changed sources

* no-mistakes(lint): Fix regression fixture types and format affected tests

* no-mistakes(ci): Fixed ci-macos-1 in test/gap-ledger.test.js:52 by realpath-normalizing the temporary root once in harness(), matching production checkoutRoots behavior. This shared correction covers all four sibling routing consumers and keeps State paths, transcript cwd, and checkout roots consistent across macOS symlinked temporary directories. Product code, evidence semantics, and assertions are unchanged. Re-traced the failing and ordinary paths before verification. Ran node --test test/gap-ledger.test.js with TMPDIR set to the value returned by getconf DARWIN_USER_TEMP_DIR: all 49 tests passed, including both reported failures and sibling routing regressions; zero failures or skips. No pipeline-control or publication actions were performed; current-head forge verification remains with the outer executor

* refactor(proposal): leave nested-session root fallback to a separate change

Keep this branch to opt-in OMP discovery: an edit whose evidence lacks a native
session is routed as on main (the unattributed evidence is left out of ownerOf),
instead of falling back to the root memory file. The fallback moves to its own PR.

* no-mistakes(review): Refresh cached proposal interaction classification from discovery

* no-mistakes(document): Clarify OMP observer counting and cache documentation

* no-mistakes(lint): Format interaction regression test

* fix(discovery): confine OMP handling to the opt-in store and leave routing to main

With includeOmp unset, plain Pi stores now list and read exactly as on main:
the two-level walk, subagent tagging, ancestor lookup, and title-header parsing
apply only to the ~/.omp store when discovery.includeOmp is on. scan --json
omits the observer fields unless a transcript shares an observer with a parent.

Fold routing goes back to main's route.ownerOf over cluster sessions; the
incomplete-attribution root fallback and its native-ownership tests move to the
separate routing change.

* no-mistakes(review): Keep plain Pi ledger slots and main routing, drop leftovers

* no-mistakes(review): Revert unrelated acpx and gitignore hunks to main

* no-mistakes(review): Restore main's gapSource body in gap ledger

---------

Co-authored-by: mdc2122 <mdc2122@users.noreply.github.com>

* chore(main): release backpass 0.1.33 (kunchenguid#207)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* fix(discovery): associate sessions in repos with separate git directories (kunchenguid#206)

* fix: associate sessions in repos with separate git directories

listWorktrees() now seeds its candidates with the normalized checkout
root before adding Git's reported paths, retaining the existing
deduplication. Regression tests cover exact and nested tier-1
association, strict-mode acceptance, exclusion of unrelated directories,
ordinary and linked checkout deduplication, and fallback outside a Git
repository.

Fixes kunchenguid#127

* fix: don't count a separate git dir as a worktree

---------

Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>

* chore(main): release backpass 0.1.34 (kunchenguid#208)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* fix(memory): stop treating thematic breaks as instruction units (kunchenguid#213)

* fix(memory): skip thematic breaks when parsing instruction units

A break-only line (`---`, `***`, `___` and CommonMark variants) outside a
fence was parsed as its own instruction unit, inflating the status count,
taking an AG id, and matching any skill body break as a 1.00 overlap.
parseMemoryUnits now drops such lines and flushes the current unit. A
dash run directly under a plain paragraph is treated as a setext H2
heading, as in CommonMark; under a list item, quote, or fence it stays a
break.

Positional AG ids shift for unchanged memory text, so cached evidence
would point at the wrong units: bump ANALYSIS_INDEX_VERSION to 4.

Fixes kunchenguid#192

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* no-mistakes(review): Keep setext headings with their content when splitting removals

* no-mistakes(review): Limit split guard to ATX and setext headings

* no-mistakes(review): Skip thematic breaks when matching a heading to its content

* no-mistakes(ci): Fixed Greptile finding ci-1, as you asked. In `parseMemoryUnits` (src/memory.js), indented code was treated as a setext heading title. Because of that, ` npm test\n---\nNext.\n` dropped the command and used it as the section name for `Next.`. Rule that must hold: a line can only be a setext title if it has at most 3 leading spaces (CommonMark). A line indented 4+ spaces or with a tab is indented code; it stays a unit, and the `---` after it is a plain thematic break. Fix: one extra condition on the setext check, `/^ {0,3}\S/.test(buffer[0])`. Checking only `buffer[0]` is enough: indented code cannot interrupt a paragraph, so if the first line is indented, the whole buffer is code. The only other place that decides what a setext title is, `isSetextTitle` in `splitRemovalHunk` (src/workspace.js), already skips any line that belongs to a unit. The indented line is now a unit, so that function picks up the fix without a separate change. I added both the space-indented example from the finding and a tab-indented one to the existing "dash run under a list item, quote, indented code, or fence is a break" test in `test/memory-units.test.js`. That test fails on the original code (7 pass, 1 fail) and passes with the fix. `pnpm run check` (lint, format, typecheck, all tests) passes: 940/940

---------

Co-authored-by: Zexuan Liu <zxliu@scs336.sim>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* fix(apply): refuse a proposal made under a different budget (kunchenguid#210)

* fix(apply): refuse a proposal made under a different budget

Apply gated the accepted subset against the budget in the current config,
not the cap the proposal was measured and gated under. A proposal made with
--budget 30000 and applied without it was judged "already over" a 10,000
cap and every addition refused; the reverse skipped the zero-sum the
proposal was held to.

Apply now compares each file an edit lands in (root and every nested file)
against the cap recorded in the proposal and, on a mismatch, refuses before
the review, naming both numbers and the remedy (re-run apply with
--budget N, or nestedBudgetTokens when it is set, or propose again). The
writer repeats the check for the accepted subset and writes nothing. A
proposal that recorded no cap is gated at the current config, as before.
The writer's nested cap now resolves through nestedBudget(), the same
helper propose uses.

Fixes kunchenguid#188

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* no-mistakes(ci): I added the test Greptile asked for; no source code changed. The bot was right that the writer's own root-budget check had no coverage. The existing root-mismatch tests in `test/apply-cli.test.js` all run the CLI, and `cmdApply` refuses before `applyDecisions` is ever called, so removing the writer's check would have left every test passing. **Invariant:** accepted edits to a file must never be gated or written under a budget different from the one the proposal was made under, and the writer (`applyDecisions`) has to enforce this by itself, not just `cmdApply`. Both files that can mismatch are now tested against the writer directly: nested files already had direct tests in `test/nested.test.js`, and the root file now has one too. **The new test** is "the writer itself refuses accepted root edits made under a different budget", in `test/apply-cli.test.js`. It builds a proposal at a 30,000-token budget, then calls `applyDecisions` directly with one edit accepted, one rejected, and the config set to 10,000. It checks that: - nothing is written, `AGENTS.md` is unchanged, and no skill directory is created; - the single failure message names the file, both numbers, and the `--budget 30000` remedy; - `rejectionsRecorded` is false and `.backpass/rejections.json` holds no entries. **Verification:** I temporarily turned the writer's check off and the new test failed; with the check back in place it passes. `pnpm run check` (lint, format, typecheck and tests) passes: 934 tests, 0 failures

---------

Co-authored-by: Zexuan Liu <zxliu@scs336.sim>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* feat(status): show each nested memory file's sampled-session reach (kunchenguid#209)

A nested file trains only on sessions whose every in-repo work path stays
under its directory, so in repos where sessions cross directories it may
never reach minGapEvidence, and a run that proposes nothing there looked
the same as a file with no gaps.

Each attribution now records, per nested file, how many sampled sessions
worked only under its directory (.backpass/nested/reach.json). The run
summary and `backpass status` print that count and warn below
minGapEvidence, when no add, rewrite, or remove edit can reach the file.
`status --json` reports it as nestedAttribution, and the proposal's nested
entries carry sampledSessions, minGapEvidence, and belowMinGapEvidence.
The count comes from deterministic attribution and is report-only.

Refs kunchenguid#191

Co-authored-by: Zexuan Liu <zxliu@scs336.sim>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* chore(main): release backpass 0.1.35 (kunchenguid#215)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* docs: clarify apply UI dependency (kunchenguid#216)

* fix(memory): recognize CLAUDE.md symlinks as pointers (kunchenguid#217)

* fix(memory): recognize Claude memory symlinks

* no-mistakes(document): Document canonical memory symlink resolution

* chore(main): release backpass 0.1.36 (kunchenguid#218)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* feat(omp): accept omp as an analysis and synthesis runner

omp (Oh My Pi) is an acpx custom agent. Its model and reasoning level now travel
over acpx: --model at session create, and the ACP `thinking` session option.

- harness-invoke: omp carries model and effort through the session path; a level
  omp does not advertise (off, auto, low, high, max) is refused before any call,
  so a level is never silently downgraded. omp is not a built-in adapter, so the
  agents.<name> replacement guard does not apply to it.
- agents: resolvedEffort gives omp no default that it cannot carry (analysis
  medium). A configured level is kept as written.
- acpx: omp's session effort key is `thinking`, so an exec fallback fails loudly.
- README: one runner line for omp.

Tests cover the transport, the refusal, the loud failure, the configured-agent
acceptance, the effort default, and the probe's openai-codex route mapping.

* no-mistakes(review): Qualify omp README ids, add omp to hints, test routes

* no-mistakes(document): Update stale effort-default comment for omp

* test(agents): use neutral providers in colliding omp ladder test

Replace literal openai/ and openrouter/ ids with neutral provider names.
The test still proves the qualified ladder id resolves to itself when
the same bare model is advertised by other providers.

---------

Co-authored-by: TNTHNGVDYNND <161745519+TNTHNGVDYNND@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Andrew Sweet <andrew.sweet@cantab.net>
Co-authored-by: firstmate <firstmate@local>
Co-authored-by: Kun Chen <3233006+kunchenguid@users.noreply.github.com>
Co-authored-by: Admin Athletics mind <dev@athleticsmind.com>
Co-authored-by: Rodrigo Díaz Jonguitud <dijongui@gmail.com>
Co-authored-by: Matt Van Horn <mvanhorn@users.noreply.github.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Kun Chen <kunchenguid@users.noreply.github.com>
Co-authored-by: Kenmege <205099287+Kenmege@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Pantoria <37699442+Pantoria@users.noreply.github.com>
Co-authored-by: zxliu <zxliu@scientechresearch.io>
Co-authored-by: Pawan <pawan161096@gmail.com>
Co-authored-by: Daniel Martinez Morales <dmmdea@hotmail.com>
Co-authored-by: mdc2122 <mdc2122@users.noreply.github.com>
Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
Co-authored-by: Zexuan Liu <zxliu@scs336.sim>
Co-authored-by: Som Samantray <92726151+SomSamantray@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Support nested AGENTS.md in monorepos (train per-directory memory files from the sessions that worked there)

2 participants