@aiwayds/dsh-topics-memory

Configuration

First-time setup belongs to /topics onboard; day-to-day tuning is /topics set — on dsh 0.1.7+ it writes the settings-page dsh-topics-memory entry (the profile patch) and every key is volatile: edits take effect immediately, no restart. Upgrading from a pre-0.1.7 install: a legacy top-level topics: section in the old settings.yaml is imported once into the new entry automatically at the next plugin boot (the audit record lands in ~/.dsh/storages/dsh-topics-memory/legacy-import.json). All keys and defaults:

Key	Default	Meaning
repo	empty (local-only)	GitHub sync repo owner/name; suggested dsh-topics-data; empty = back to local-only
autoInject	true	Per-turn injection master switch
injectDedup	true	Session-level injection dedup: topics already injected in this session are not re-injected (registry cleared at session end; budget-dropped topics stay injectable; deduped topK slots are NOT backfilled) — ADR 0012
suppressEcho	true	Distill-echo suppression: topics distilled from the CURRENT session's own turns are not injected back into it (provenance rides the observations log sessionId → distilledInto)
topK	4	Max topics injected per turn
perTopicBudget	300	Per-topic digest token budget
totalBudget	1500	Total injection budget per turn
matchThreshold	0.3	Hit threshold; tune from /topics stats near-miss evidence
tagBoost	0.15	Additive boost per tag hit (total cap across hits equals this value)
injectMode	pointer	Injection shape: pointer (light pointers, ≤80 tok each, topic_open pulls the full text; total budget capped at 600) / digest (full digest rendering, per-topic 300 / total 1500)
qualityLane	sampled	Slow quality lane: off / sampled (1/3 of turns) / always; produced at turn/end, consumed by the next injection (consume-once), never for subagent sessions
graphDepth	2	depends graph walk depth (0 disables)
recencyWindowDays	7	Recency bonus window (+0.2)
autoObserve	true	Capture atomic observations every turn
includeSubagents	false	Whether injection and observation also engage subagent sessions (ADR 0011; off by default since 0.7.0); off skips them entirely
observationMaxChars	2000	Per-side per-turn observation truncation
distillProvider / distillModel	empty (distill off)	Distill lane model route; both must be set to enable. With a UI, /topics set distill-provider / distill-model without a value opens a picker panel (provider list → that provider's model catalog); a mixed provider model / provider/model value for distill-model splits into both keys
distillEveryTurns	5	Distill every N turns of a long session
distillOnSessionEnd	true	Distill once when a session ends
distillBatchSize	40	Observations per distill model call. On an output-limit (max-tokens) failure the batch halves automatically (floor 5) and retries — a failing batch can no longer livelock the backlog; the shrink persists until reload or a config change. Note: /topics set distillBatchSize back to the same value does not reset the shrink — set a different value or reload the plugin
distillMaxModelCalls	8	Max model calls per distill run, including the one corrective retry for ops echoing no valid observed_ids (the run stalls when the budget can't fit it). Batches already distilled keep their marks when the budget stops the run (partial progress), recorded as partial: … in the distill state
consolidateCadence	daily	Consolidation-lane cadence: daily/3d/7d/off. At session start the plugin checks the last consolidation time (meta/consolidate-state.json) and, when due, runs the LLM gardener in the background (reusing the distill model route): local lexical clustering only feeds near-look-alike candidate clusters; four op kinds merge/promote/deprecate/refresh, out-of-scope ops are dropped; a failed call never advances the stamp, so the next start retries
deprecatedTtlDays	15	Deprecated topics older than N days are dropped at session start (local rule, no model; each drop is its own git commit — history stays recoverable); 0 disables the sweep
usageBoost	0.15	Usage boost (ADR 0015): topics injected/opened in the last 30 days score higher at retrieval (gate-scoped, capped at 0.2, never granted to zero-lexical candidates, structural gate not waived); 0 disables
zeroOpenDecay	false	Zero-open decay: halve the retrieval score of zero-open high-frequency slugs before the threshold cut (see the section below)
pushDebounceSeconds	45	GitHub-mode debounced push interval
jevEnabled	false	System One decision layer master switch (experimental): false = zero behavior change — see the Jev decision layer section above
jevBackend	zen	Decision endpoint: zen (free) / native / openrouter
jevModel	empty (per backend: jev-1.13-free / jev-1.13.0 / typesafe/jev-1.13)	Version-pinned decision model; upgrading is an explicit action
jevTimeoutMs	3000	Per-request decision timeout; single attempt, no retry
jevSecretFile	none	External secret list for the outbound secret gate; re-read when hot-changed
jevLayaFallback	false	Local laya pace-maker: fire a parallel laya call on every jev request; its answer takes over (degraded, relative-ranking only) when the primary fails
jevLayaUrl	http://127.0.0.1:8000/v1/systemone	laya-serve endpoint for the pace-maker

Zero-open decay (default off)

zeroOpenDecay: true halves the hit score of slugs the stats panel lists as 零打开高频 — truly rendered ≥5 injections in the rolling 30 days with zero topic_open events ever. The decay applies at score finalization, before the match threshold, so a decayed candidate that drops below it lands in nearMisses carrying the zero-open-decay reason — the injection log shows why it stopped ranking. Motivation: 24 high-frequency topics held the injection leaderboard with zero opens while score and opens proved only weakly correlated (the 0.3 threshold stays untouched). Default false means zero behavior change; unreadable injection/open logs skip the decay (fail-open). Toggle live with /topics set zero-open-decay on|off.