17. Configuration

No hardcoded tunables anywhere the harness can configure them. The table below is the complete host-plane list (the config schema in src/index.ts); the comment header of the shipped cordis.patch.yml (inside node_modules/@fractalsynapse/dsh-chapters/) mirrors the schema and is the in-repo reference, but this page is kept current with the code.

Two planes. config (below) is the host plane — your profile’s plugin entry. The compaction engine lives on the realm plane: the chapters agent-preset row (the preset-chapters declaration row at the foot of the bundle’s cordis.patch.yml, mounted by src/preset.ts), which owns thresholdRatio, retainRatio/retainTokens, maxTokens, compactionRetries/maxOverflowRetries, summarizationProvider/summarizationModel (legacy override), enrichmentTrigger (afterPush|idle|both|manual), elicitedPlot, and toolResultArtifactTokens. Shared names behave per plane — design record §12.

The #50 rule keys below are declared on the host plane only; the three the compaction checkpoint also needs (projectRulesDisplayCap, sessionRulesDisplayCap, rulesRelevance) are mirrored into the engine’s chaptersConfig the way elicitedPlot and enrichmentEnabled already are — one declaration read by both planes, so a head rebuilt by either path selects the same rules.

Key	Default	Meaning
harnessId	hostname	this machine’s identity: commit author, curation partition, rule authorship
knowledgeRemote	''	pre-bind an upstream (else /chapters-link)
knowledgeProvider	auto	transport policy: auto dispatches on the remote scheme (treedx+ → TreeDX, else git); an explicit git/treedx must match the scheme or the link refuses
projectKeyOverride	''	wins over the remote-derived project key
syncDebounceMs	30000	coalesce window for post-archive pushes
treedxFetchTimeoutMs / treedxWorkspaceTtlSeconds / treedxLeaseRetries / treedxLeaseRetryDelayMs	15000 / 900 / 3 / 1000	TreeDX transport tunables (fetch timeout, workspace TTL, lease retry count and delay)
enrichmentEnabled	true	false makes the whole P2 ladder a pure no-op — corpus stays fully working
enrichmentModel	'' (conversation route)	the annotator override
enrichmentMaxTokens	4000	one auxiliary completion budget (annotation, mined rules, the conflicts report). It has to cover reasoning tokens, not just the visible answer — measured mining usage ran 1.1K–3.2K per chapter. The route used here ignores max_tokens, so this is headroom for providers that don’t rather than a live limit
enrichmentIdleMs / enrichmentBatchCap	60000 / 5	idle trigger; chapters per drain
vocabApply	false	shadow (default) vs git-visible topic curation
vocabCoMin / vocabOverlapMin	3 / 0.5	trust thresholds for emergent vocabulary
coreRulesBudgetTokens	1200	the CORE RULES block in notices — refuses with numbers, never clips
rulesCoreBonus	0.15	per-machine rank bonus for core-approved rules
sessionRulesEnabled	true	deterministic ADD RULE: capture into the local session tier — zero model tokens, never staged to the pool
rulesMinedPerChapter	3	cap on mined rules per chapter; the cap counts rules written, so a duplicate cannot consume a chapter’s budget, and 3 is the ceiling on what any one archived chapter can teach the session
rulesMarkerPrefix	ADD RULE:	the capture marker; matched at line start in human-authored messages only, so prose that merely mentions it is not a rule
rulesEnrichment	true	model-derived capture: the enrichment pass mines standing orders from each archived chapter, idle-batched, never per-turn
rulesAutoActivate	true	whether a mined rule governs on arrival. It does by default — inferring what nobody said is the tier’s purpose (#50, Amendment 2026-10-21). Off ⇒ mined rules are written reviewOnly and stay inert until you approve
rulesModel	''	'provider/model' for a dedicated mining model: one extra bounded call per chapter, and its answer replaces (never supplements) the annotation pass’s rules. '' ⇒ mine inside the enrichment call, at no extra cost. Runtime: /chapters-rule model p/m
projectRulesDisplayCap / sessionRulesDisplayCap	12 / 12	per-tier count caps on rules rendered into a rebuilt head — anything dropped is announced by count, never silently clipped
rulesRelevance	true	topic-overlap selection on (search.rankEntries, the same scorer as chapters_search); off ⇒ deterministic category→id order
promptModel	''	the model that defines the next work (§28). Empty ⇒ the feature does not exist — zero calls, zero surface. Setting it (here, or at runtime with /chapters-prompt model p/m) is the enable: one knob, no second boolean to drift
promptContextChars	derived	budget for the suggestion input; model-derived by default — contextWindow × promptContextRatio at 4 chars/token when the prompt model is set via /chapters-prompt model, this key is the fallback when no window resolves, and the command stores an explicit pair beside the model when given numbers (#188)
promptContextRatio	0.5	share of the model’s context window the derived budget takes — one rule, budgets 30× apart between a 32K and a 1M model
promptMaxChars	4096	cap on the recommendation (raised from 240 by #188 — a cap that cannot carry a paragraph cannot carry an instruction; settable beside the model). An over-cap answer is retried once with the cap restated, then refused — never truncated
—	—	all three model keys take a provider/model route — /chapters-models lists the exact identifiers this host accepts, their context windows, and each key’s current value (#173)
—	—	/chapters-rule conflicts has no key of its own: it runs one bounded stream call on the conversation route the session already has, report-only, never edits a rule. No route ⇒ the verb refuses honestly
continuationBudgetRatio	0.25	TOC + note share of the window remaining after the header
chapterTokenTarget / chapterLimit	8000 / 8000	soft per-chapter size / hard split
mergeThreshold	0.3	turn→chapter topic-merge τ (archive-time only)
subagentAutoSave	true	archive a subagent’s completed turns into searchable chapters when its run settles (every settlement of a continuable child; the watermark makes repeats save only what is new). Save-only — creates no session, compacts nothing; enrichment is queued, never run per child. false unregisters the behavior entirely
planInlineChars	900	verbatim plan characters printed in a chapter’s Plans block — the plan’s full text is always an artifact, so this is display economy, never a data decision (0 = pointer only)
plansInNotice	true	carry the plan block into a child’s notice; renders zero bytes when the path holds no plan, so tape neutrality is the default
planSeedMode	true	seed plan/mode {active:true} into a continuation child when the archived span was still in plan mode, so the gate survives the boundary (measured: kernel accepts the two-event seed, the host’s own projection folds it). Never for a branch, never after an approval — approving a plan exits plan mode
plansInCheckpoint	true	carry the operative plan into every in-place compaction checkpoint, the way the plot is carried — because shadowing can hide the chapter that holds it
planCheckpointChars	1600	how much plan text a checkpoint carries; the archive keeps all of it, and a truncated carry says so in the text
retrievalHints	true	put `## Loudness (issue #61)

key	default	what it changes
diagnosticsLog	unset → <store-root>/diagnostics.log	where refusals, deferred finalizations and door failures are appended. DSH_CHAPTERS_ENGINE_ERRORS overrides it; the literal off disables file logging (the profile logger still gets them). Before this, an unset variable meant no trace on disk — and three real compaction failures left none.
warningsDisplayCap	6	how many warning TEXTS a door prints before switching to …and M more. Counts alone hid 86 warnings across six saves, including the one that named a hollow chapter.
segmentCandidateCapTokens	4000	the ceiling on the oversized-result candidate list chapters_segment returns. Ranked by size, and the response always names what was withheld.

Finding things in the child's notice — the three lookup verbs (read, chapters_artifact, chapters_search`). Off ⇒ the notice is byte-identical to the pre-hint shape |

| callArgArtifactTokens | 0 (off) | above this, a tool-call argument is artifact-backed and the inline text is a bounded head. Off by default on purpose: switching it on changes chapter bytes wherever a write carried a whole file, which means your model tapes need re-recording | | toolResultDeferFloorTokens | 200 | archive-time inline/defer floor (realm toolResultArtifactTokens is the arrival floor — keep it above your chunk size: below it, stubs hide compaction and thresholds silently stop firing — measured) | | artifactStoreRoot | .dsh-chapters | workspace store (must be read-reachable) | | noticeIndexCap | 6 | how many chapter lines the continuation notice lists; 0 lists every one. Past the cap the notice says so and points at chapters_search — nothing is dropped silently. The point is SIZE stability: an uncapped list grows ~120 chars per chapter, so the surface crosses the compaction trigger at a different turn on each run, which makes a boundary-sitting journey impossible to capture as a replayable tape (#56) | | installChaptersPreset | true | inert since 0.1.18 — the preset is the preset-chapters row in the bundle patch; disabled: true on that row is the switch. Still accepted so an older profile patch parses | | compactionThresholdRatio | 0.9 | inert since 0.1.18 — it is thresholdRatio on the preset’s chapters-engine row now (see §17 notes on the realm plane) | | fallbackPreset | chapters | mounted for children when the caller’s preset can’t be resolved | | searchDefaultMaxTokens | 400 | chapters_search result-packing budget |

Read the row values, change what you own, restart to reload. Next: 18. The continuation notice, annotated.