dsh-rich-indexing

Tiered, index-aware compaction for DeepSeek Harness. Takes over the compaction service from compaction-basic and replaces the single 0.8 cliff with a graduated ladder, escalating consolidation laws, a deterministic keyword index folded into every checkpoint, and a summarization model chain.

What changes vs stock compaction

	stock basic	rich-indexing
Trigger	one zone (thresholdRatio: 0.8)	tier ladder, default 30% / 50% / 70% / 90%
Summarization	one fixed law	four laws, escalating per tier: gentle → standard → consolidating → maximum
Terminology	model's choice	deterministic keyword index (paths, recurring identifiers, versions) merged into every checkpoint; prior index fed back for accumulation
Summarizer	one provider/model pair	primary + up to 3 fallbacks, each with its own reasoningEffort; last resort = the conversation's own route
Everything else	—	inherited unchanged: lock, head-anchored checkpoint merge, tool-pair balance, overflow recovery, /compact, token-meter pricing

The checkpoint structure stays the stock 8 sections (plus Problem-Solving Notes and verbatim User Messages at the gentle/standard tiers), so downstream consumers see the same durable <compacted-summary> shape.

Takeover semantics (read this)

Cordis allows exactly one provider per service key, so this plugin disables compaction-basic with one managed line in the profile's own cordis.patch.yml:

# managed:rich-indexing
- id: compaction-basic
  disabled: true

• Enable: written automatically at plugin start; the web profile's live patch reload swaps the engines without a restart (otherwise after the next restart — the panel shows pending restart until then).
• Disable (toggle the plugin off): the managed line is removed automatically; stock compaction returns live. Compaction is never left dead by a toggle.
• Uninstall (not toggle): has no teardown owner — delete the managed block above from ~/.dsh/profiles/web/cordis.patch.yml, or press Release in the panel before uninstalling.
• Release (POST /api/rich-indexing/release): removes the line and self-disables the plugin; stock behavior returns.

Install

dsh plugin --profile web install ./dsh-rich-indexing

Requires the profile to include the token-meter (any deployment running stock compaction already does). Peer imports (@deepseek-ai/dsh-compaction-*, @deepseek-ai/dsh-llm, @deepseek-ai/dsh-settings, @deepseek-ai/schemastery) resolve through the profile module fallback from the installation closure.

Configuration

Settings → Plugins → rich-indexing (persisted in ~/.dsh/settings.yaml under the rich-indexing namespace, live-applied), or the equivalent entry config:

enabled: true
tiers:                      # strictly increasing ratios; retain < ratio per tier
  - { ratio: 0.3, retainRatio: 0.12, law: gentle }
  - { ratio: 0.5, retainRatio: 0.1,  law: standard }
  - { ratio: 0.7, retainRatio: 0.08, law: consolidating }
  - { ratio: 0.9, retainRatio: 0.05, law: maximum }
models:                     # primary first; up to 4 routes total
  - { provider: deepseek-official, model: deepseek-reasoner, reasoningEffort: high }
  - { provider: deepseek-official, model: deepseek-v4-flash, reasoningEffort: default }
maxTokens: 8192

Empty models means "use the conversation's own route" (stock parity). A crossing of tier N runs one compaction at tier N's law and retention; the next crossing fires the next tier. context-overflow always runs the maximum law.

Status API

• GET /api/rich-indexing/state?sessionId=… — pressure fraction vs the tier ladder, tier pointer, last compaction facts, model-chain health, takeover state, resolved config.
• POST /api/rich-indexing/compact {sessionId} — the button form of /compact.
• POST /api/rich-indexing/release — remove the takeover line and self-disable.

Loopback + same-origin guarded, like every plugin-family route.

Client (module sub-tab + settings card)

The client half registers two slot contributions (no separate build — same __ModuleLoader__ facade format as the rest of the plugin family):

• Compaction — a conversation.view module sub-tab: live pressure vs the tier ladder, tier pointer, last checkpoint facts, model-chain health, Compact-now and Release actions, 5s auto-refresh.
• rich-indexing — a Settings → Plugins card keyed to the settings namespace: enabled toggle, maxTokens, the tier table (ratio / keep / law), and the model chain editor — provider, model, and effort per route, with lists from the host's model catalog. Saves are revision-fenced (remote.settings.mutate), live-applied, and persisted to ~/.dsh/settings.yaml.

Deployment note

pnpm add file: snapshots the package — it is not a live link. After changing this repo, re-sync the installed copy before restarting the service:

rsync -a --delete --exclude .git --exclude node_modules \
  /home/sysadmin/.dsh/plugins/dsh-rich-indexing/ \
  /home/sysadmin/.dsh/profiles/web/node_modules/dsh-rich-indexing/
systemctl restart cli-dsh-web.service

Testing

npm test        # 21 tests: pure core (tiering, index, takeover line, laws) + engine dispatch probe
                # (the probe self-skips outside a profile tree — it needs the installation closure)

Design

See DESIGN.md — the architecture copycats DSH's compaction capability seam (engine subclass, sanctioned hook points) and adopts the applicable strategies from the OpenSDD context-compactor (structured template with recall sections, analysis scratchpad, no-recursion, always-recoverable summarizer).

License

MIT