Observation
When capture followthrough exceeds its budget, the raw thought is committed and stays readable through inspect — the trapdoor working as designed. The read model does not survive it.
Reproduction, against main:
M=$(mktemp -d); git -C "$M" init -q
for i in A B C; do
THINK_REPO_DIR=$M THINK_CAPTURE_FOLLOWTHROUGH_TIMEOUT_MS=120000 \
node bin/think.js "healthy $i"
done
# recent -> 3 entries, stats total -> 3 (correct)
THINK_REPO_DIR=$M THINK_CAPTURE_FOLLOWTHROUGH_TIMEOUT_MS=1 \
node bin/think.js "DEFERRED D"
# recent -> 1 entry (only DEFERRED D), stats total -> 1
# A, B and C have vanished from both surfaces
THINK_REPO_DIR=$M THINK_CAPTURE_FOLLOWTHROUGH_TIMEOUT_MS=120000 \
node bin/think.js "healthy E"
# recent -> E, C, B, A stats total -> 4
# D is now gone permanently; the total should be 5A single deferral does three things:
- Hides every older capture while the deferred entry is newest.
- Loses the deferred capture permanently from both surfaces once a later healthy capture rebuilds the read model.
- Leaves recall inconsistent. In this reproduction
remember returned zero matches; the archived mind in test/fixtures/cas/readme-smoke-mind.json returns one for its own deferred capture. That disagreement is itself the defect.
Not introduced by the followthrough-budget work
Confirmed on origin/main at 24e2613 by temporarily forcing the then-hardcoded CAPTURE_FOLLOWTHROUGH_TIMEOUT_MS to 1: same result. Exposing the budget as THINK_CAPTURE_FOLLOWTHROUGH_TIMEOUT_MS only made the state easy to reach deliberately.
Likely mechanism
src/store/read-model.js writes fastCaptureRecordsJson as a single-element array during the raw save; the full recentCaptureRecordsJson list is reconciled during followthrough. Abandon the followthrough and the single-record fast index is what recent and stats read. The later rebuild sources from the graph, where the deferred capture was never linked (canonicalThought.stored is false), so it is omitted.
Why this is bad
Deferral is reachable in normal use — a cold repository with Git fsmonitor enabled can exceed the 6s budget on its first write, which is what made an MCP acceptance assertion flaky. An agent that captures a decision, receives the documented "not a failure" warning, and moves on has silently dropped that thought out of every surface it would later search. For a memory product this is the worst available failure mode: silent, and indistinguishable from never having captured.
Suggested direction
Either write a complete fast index during the raw save so it is never a partial view, or have the read model treat a single-record fast index as unreconciled and fall back to a full scan. The rebuild should also include raw entries whose canonical thought was never stored.
References
- Backlog card with the same repro:
docs/method/backlog/bad-code/CORE_deferred-capture-corrupts-the-recent-read-model.md
- Frozen regression fixture:
test/acceptance/readme-smoke-mind-fixture.test.js pins the deferred state, which cannot be produced on demand
Observation
When capture followthrough exceeds its budget, the raw thought is committed and stays readable through
inspect— the trapdoor working as designed. The read model does not survive it.Reproduction, against
main:A single deferral does three things:
rememberreturned zero matches; the archived mind intest/fixtures/cas/readme-smoke-mind.jsonreturns one for its own deferred capture. That disagreement is itself the defect.Not introduced by the followthrough-budget work
Confirmed on
origin/mainat24e2613by temporarily forcing the then-hardcodedCAPTURE_FOLLOWTHROUGH_TIMEOUT_MSto1: same result. Exposing the budget asTHINK_CAPTURE_FOLLOWTHROUGH_TIMEOUT_MSonly made the state easy to reach deliberately.Likely mechanism
src/store/read-model.jswritesfastCaptureRecordsJsonas a single-element array during the raw save; the fullrecentCaptureRecordsJsonlist is reconciled during followthrough. Abandon the followthrough and the single-record fast index is whatrecentandstatsread. The later rebuild sources from the graph, where the deferred capture was never linked (canonicalThought.storedisfalse), so it is omitted.Why this is bad
Deferral is reachable in normal use — a cold repository with Git fsmonitor enabled can exceed the 6s budget on its first write, which is what made an MCP acceptance assertion flaky. An agent that captures a decision, receives the documented "not a failure" warning, and moves on has silently dropped that thought out of every surface it would later search. For a memory product this is the worst available failure mode: silent, and indistinguishable from never having captured.
Suggested direction
Either write a complete fast index during the raw save so it is never a partial view, or have the read model treat a single-record fast index as unreconciled and fall back to a full scan. The rebuild should also include raw entries whose canonical thought was never stored.
References
docs/method/backlog/bad-code/CORE_deferred-capture-corrupts-the-recent-read-model.mdtest/acceptance/readme-smoke-mind-fixture.test.jspins the deferred state, which cannot be produced on demand