@rasensio/aidlc-content 1.74.0 → 1.76.0

package/docs/knowledge-graph.md

@@ -150,3 +150,27 @@ deliberate:
 
 The same shape guards the coverage cache in `src/traceability/cache.ts`, which is where the pattern
 came from.
+
+## When a source file breaks
+
+A merge that leaves conflict markers in `modules.yaml`, or a hand edit that makes one document
+unparseable, takes that document out of the index. That part is right: the YAML is the source of
+truth, and keeping the old rows would be serving stale facts again. What was wrong is that nobody
+was told. The parser named the file in a warning that nothing printed, and when no entity was left
+the `# Codebase Knowledge` block simply disappeared from `CLAUDE.md` — a shrinking diff in a
+generated file, the exact shape of thing six instances had already filed as noise.
+
+So `aidlc update` and `aidlc init` now print `knowledge graph: <file>: …` for every broken source,
+and say so when the block is left out. Three choices behind that:
+
+- **The warning is read from the files on every run**, not only on the run that rebuilt the index.
+  Whichever command opens the index first after a merge does the rebuild; if that was `aidlc gate`,
+  the warning used to be gone before `aidlc update` ever saw it. That costs one extra parse on those
+  two commands — 70–160 ms here — and nothing anywhere else.
+- **A document the index cannot store is skipped and named, not thrown.** A `description:` holding a
+  map used to fail the whole rebuild with "cannot be bound to SQLite parameter 4", which names
+  neither the file nor the field. The rule is what the index actually refuses, measured against it,
+  rather than the stricter rules `aidlc knowledge add` enforces — applying those on read would drop
+  facts that index fine today.
+- **`compact` is the exception.** It rewrites the YAML from what it read, so skipping a document there
+  would delete it from the source of truth rather than from a cache. It still reads everything.

package/docs/roadmap.md

@@ -167,7 +167,9 @@ Each item goes through these steps:
    pass is what held a whole `--all` run on one item for five hours. An agent that exits with an
    error within 30 seconds of launch parks the item with `agent failed to launch (exit <code>):`
    and the first line it printed, so a flag the agent no longer accepts names itself instead of
-   reading as a failed build.
+   reading as a failed build. An agent command that cannot be started at all — not installed, or
+   not on the background run's `PATH` — parks the item at the next pass with
+   `agent failed to launch: spawn <command> ENOENT`, and the run carries on.
 
    **The agent reports a typed result.** With one of the Claude Code or Codex presets, autopilot
    asks the agent how it ended: Claude Code runs with `--output-format json` and an outcome schema,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rasensio/aidlc-content",
-  "version": "1.74.0",
+  "version": "1.76.0",
   "type": "module",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",