Brain Store Fallback

Brain Store Fallback

Use this skill whenever a brain_store call might not have durably written to BrainLayer.

Wave evidence: 72h-mine found BrainLayer DB-busy and MCP transport closures as the #2 severity-10 ecosystem pattern, with silent brain_store loss across six or more recent sessions. See pain-points/_consolidated.md.

Core Rule

If brain_store does not return a concrete chunk_id, do not report that the content was stored in BrainLayer.

Instead, write the exact original content to:

docs.local/decisions/<YYYY-MM-DD>-<slug>.md

Then report:

Queued to docs.local/decisions/<file> pending BrainLayer recovery.

Activation Signals

Activate on any Tier 1 hard failure:

• Transport closed
• -32000: Connection closed
• apsw.ReadOnlyError
• database is locked
• Python exception text in the MCP error channel
• brain_store returns null, undefined, or no chunk_id
• Probe result includes TOOL_RETURNED_TEXT=no

Activate on any Tier 2 soft queue signal:

• chunk_id: null
• Response says queued, queue depth, deferred, or similar without a durable chunk id
• The call takes longer than 25 seconds

Treat Tier 3 signals as advisory:

• The latest brain_search also failed with a transport error
• Other BrainLayer tools are returning connection failures

Tier 3 means be ready to fall back on the next brain_store; it does not by itself replace the primary brain_store attempt.

Fallback Contract

The fallback file is a local replay queue entry, not a BrainLayer ack.

Write YAML frontmatter followed by the exact original content:

---
intended_brain_store: true
importance: <1-10>
tags: [<original tags>]
timestamp: <ISO 8601>
reason: <transport_closed | db_busy | queued_null_chunk | readonly | unknown>
retry_attempted: false
chunk_id:
---
<original brain_store content, byte-for-byte>

Preserve all four brain_store semantics:

• Content
• Importance
• Tags
• Timestamp

Do not add fallback annotations to the body. Do not reformat the body. Do not drop tags or importance.

Filename Rule

Use the original timestamp date plus a deterministic slug:

docs.local/decisions/<YYYY-MM-DD>-<slug>.md

If --slug is supplied, sanitize it:

• lowercase
• spaces to hyphens
• keep only letters, numbers, and hyphens
• collapse duplicate hyphens

If no slug is supplied, derive one from the content:

<first-8-tokens>-<hash-prefix>

Where:

• first-8-tokens: split the content on whitespace, trim leading and trailing punctuation from each token, lowercase each token, keep only letters, numbers, and hyphens, drop empty tokens, take the first eight, then join with hyphens.
• hash-prefix: first eight characters of the lowercase hex-encoded SHA-256 hash of the original byte-for-byte content body.

The same content with the same original timestamp must resolve to the same filename. This prevents retry pollution.

Immediate Workflow

1. Call brain_store first.
2. Inspect the result.
3. If chunk_id is present and non-empty, report the BrainLayer chunk id.
4. If any activation signal appears, write the fallback file.
5. Retry brain_store once after a short pause.
6. If retry succeeds, update fallback frontmatter with:
   • retry_attempted: true
   • chunk_id: <returned id>
7. If retry fails, update fallback frontmatter with:
   • retry_attempted: true
   • chunk_id:
8. Continue the session without retry-storming.

One retry per call site is the maximum. More retries make DB contention worse.

Reporting Contract

Use honest status language:

• Stored in BrainLayer: <chunk_id> when a concrete chunk id exists.
• Queued to docs.local/decisions/<file> pending BrainLayer recovery when the fallback file exists but no chunk id exists.
• BrainLayer store failed and fallback write failed only if both paths failed.

Never say stored, saved to BrainLayer, or BrainLayer updated when only the fallback file exists.

Replay

Boot-time replay lives in workflows/replay.md.

At the start of any session that expects to call brain_store, scan docs.local/decisions/*.md for files with:

• intended_brain_store: true
• missing or empty chunk_id

Replay each pending file through brain_store.

On success, update frontmatter in place with the returned chunk_id. Leave the file where it is. Do not delete it. Do not auto-move it to _resolved/.

There is no replay cap for now because the expected queue is small. If a boot scan finds more than 100 pending files, stop and add a replay budget before processing them.

Hook

The frontmatter hooks: block is forward-compatible with per-skill hook support. It covers hard transport/database failures, null chunk ids, common queue text, TOOL_RETURNED_TEXT=no, and slow calls when the runtime exposes elapsed time. If the runtime does not invoke it automatically, or does not expose one of those result fields, apply the manual workflow above after every brain_store call.

Anti-Patterns

Silent Skip

If brain_store returns a queue ack, null chunk id, slow response, or transport error and you do nothing, you are reproducing the severity-10 failure.

Retry Storm

Do not loop until BrainLayer recovers. One retry, then fallback queue.

Body Drift

Do not add notes such as (fallback) to the body. Replay deduplication depends on the original content remaining byte-for-byte identical.

Missing Metadata

Do not omit importance, tags, or timestamp. Replay without these loses the search value of the original store.

False Ack

The fallback file is not proof BrainLayer has the content. Report the fallback path, not a fake store success.

Composition

• /never-fabricate enforces the status language: fallback files are real, but they are not BrainLayer storage.
• /frustration-capture can write the user's verbatim feedback into this fallback path when BrainLayer is unavailable.
• /session-handoff should list pending fallback entries so the next session can replay them.
• /orc can remind workers to invoke this skill when BrainLayer write cadence drops under load.

Evals

Fixtures live in evals/fixtures/.

Done Criteria

• docs.local/decisions/ is the fallback queue path.
• Fallback files preserve content, importance, tags, and timestamp.
• Replay scans pending files and updates frontmatter in place.
• Hook metadata documents the automatic trigger for future runtime support.
• User-facing output never claims BrainLayer storage without a real chunk_id.
• Idempotency maps the same content and timestamp to the same filename.