[cryptotavares]

Summary: Remove the knowledge store and its per-tool step recording. Keep session metadata as an explicit consumer write, separate from tool execution and observation collection.

Stack context: mm CLI v2, Phase 1, P1-04, decision D6. Base: #59; next: #61. Review this delta against #59.

Problem

The shared HTTP executor awaits a knowledge-store write after each tool execution with a session ID. It also records execution failures. The store persists tool inputs, outcomes, observations, and artifact references under test-artifacts/llm-knowledge/<sessionId>/steps/.

Four knowledge_* tools read that history. Decision D6 removes this subsystem from the core. Session metadata and HTTP observations still need independent homes after the store disappears.

Solution

Delete the store, its recording hooks, retrieval tools, schemas, CLI commands, and public exports. Remove knowledgeStore from ServerConfig and ToolContext, so tool execution no longer depends on a history service.

Keep two boundaries explicit:

• The daemon collects observations and builds responses through the existing execution path.
• Consumers call the exported writeSessionMetadata(artifactsDir, metadata) when they persist launch metadata.

flowchart TD
    A[Tool request] --> B[Validate and run tool through queue]
    B --> C[Collect observations under existing policy]
    C --> D[Build timed HTTP response]
    E[Consumer launch metadata] --> F[Validate session ID in writeSessionMetadata]
    F -->|valid| G[Write session.json beneath artifactsDir]
    G --> I[Return absolute file path]
    F -->|invalid| H[Reject; no metadata write]

    classDef new fill:#fff3cd,stroke:#d39e00
    classDef stop fill:#f8d7da,stroke:#c82333
    class F,G,I new
    class H stop

Loading

Yellow boxes are new. Red boxes stop the metadata write.

The metadata writer creates missing directories, overwrites session.json, and writes formatted JSON with a trailing newline. Relative artifact directories resolve against the process working directory. Invalid session IDs reject with Invalid session id: "<id>" before directory creation.

Move StepRecordObservation and createDefaultObservation to src/tools/types/observation.ts. Move SessionMetadata to src/tools/types/session.ts, without goal or flowTags. Keep free-form tags.

What changes

Condition or boundary	Before	After
Tool execution with a session ID	Awaits a step-record write after execution	No knowledge-store recording on success or execution failure
Tool registry and schemas	Include four knowledge_* tools	Remove those tools; registry has 43 entries
mm knowledge-*	Routes to history tools	Unknown command; exit 1
Launch --goal or --flow-tags	Adds knowledge metadata	Warns about the unknown flag; omits its field
Persist session metadata	Store writes under llm-knowledge/<id>/	Consumer calls the writer for <artifactsDir>/<id>/session.json
Observation collection and response policy	Collects beyond tools that return observations	Same policy; later stack work limits collection

The CLI uses its existing unknown-command message:

Error: unknown command 'knowledge-last'. Run 'mm --help' for usage.

Removed launch flags use the existing warning path, such as Warning: unknown launch flag '--goal'. The warning alone does not reject launch. The protected sections below list all breaking exports and consumer migration details.

Decisions for reviewers

Decision	Reason
Remove the subsystem under D6 / P1-04	A compatibility wrapper would retain the history dependency and its public surface. Consumers migrate explicitly.
Let consumers persist session metadata	The writer accepts their artifact directory without constructing a store or running inside each tool request.
Keep observation behavior in this PR	Read-only and discovery observations still feed the next response's diff. Changing collection belongs to later stack work.
Keep StepRecordObservation and createDefaultObservation names	Their implementations remain useful outside persistence. Moving them avoids an additional public rename.

Review guide

Read the files in this order:

#	File	What to check
1	src/server/create-server.ts, src/types/http.ts	Recording disappears from success and failure paths; observations and timing remain.
2	src/server/session-metadata.ts, src/tools/types/session.ts	Consumer-owned persistence, path validation, and retained metadata fields.
3	src/tools/types/observation.ts, src/index.ts, src/tools/types/index.ts	Moved observation contracts and public exports.
4	src/tools/registry.ts, src/validation/schemas.ts, src/cli/mm.ts	Removed tools, launch fields, commands, and help text.
5	src/server/session-metadata.test.ts, src/tools/types/observation.test.ts, CLI/server/registry tests	Metadata writes and rejection cases, moved defaults, command removal, and retained observation policy.
6	README.md, SKILL.md	Updated architecture, metadata path, command inventory, and launch options.

Testing

• Fresh check at remote head 2f10dd5ac47198f9259f24de061ee2b0bd0209a8: CI=true yarn vitest run --maxWorkers=2 --minWorkers=1 --coverage.enabled=false passes: 46 files, 1,397 tests.
• yarn lint:eslint, yarn prettier README.md SKILL.md --check, and yarn tsc --noEmit --project tsconfig.build.json pass.
• Standalone yarn tsc --noEmit --project tsconfig.test.json fails on existing mobile-driver, network-mock, and screenshot test typing issues. Vitest reports no type errors, so its result does not replace this standalone check.
• Metadata tests cover the destination path, missing directories, overwrite, and unsafe session IDs. CLI tests cover removed commands and launch flags.
• The Proof and Unit tests sections below retain the original historical evidence. These fresh checks do not repeat live consumer validation.

Breaking changes

• Removed public exports:
  • Store: KnowledgeStore, KnowledgeStoreConfig, createKnowledgeStore, setKnowledgeStore, getKnowledgeStore, hasKnowledgeStore, knowledgeStore
  • Tokenization: tokenize, tokenizeIdentifier, expandWithSynonyms
  • Tools and schemas: the knowledge*Tool functions, knowledge*InputSchema, knowledgeScopeSchema, knowledgeFiltersSchema, Knowledge*Z
  • Constants and types: FLOW_TAGS, FlowTag, STEP_LABELS, StepLabel, the StepRecord* types except StepRecordObservation, the PriorKnowledge*, Knowledge*, RecipeStep and SessionSummary types
  • Other: ErrorCodes.MM_KNOWLEDGE_ERROR, extractTargetFromInput, extractScreenshotInfo, extractToolOutcome
• ToolName no longer includes knowledge_*.
• ServerConfig.knowledgeStore and ToolContext.knowledgeStore are removed.
• Launch input no longer accepts goal/flowTags, and SessionMetadata no longer has them. mm launch --goal now gives the unknown-flag warning.
• mm knowledge-* commands are removed (unknown command, exit 1).
• Path change: session metadata moves from test-artifacts/llm-knowledge/<id>/session.json to <artifactsDir>/<id>/session.json (test-artifacts/<id>/session.json with the consumers' artifactsDir: 'test-artifacts'). Consumers call writeSessionMetadata() themselves.

Consumer impact

Consumer	Change needed	Done on branch
Extension test/e2e/playwright/llm-workflow/	Drop KnowledgeStore/setKnowledgeStore from daemon.ts; metamask-provider.ts calls writeSessionMetadata(artifactsDir, …); no goal/flowTags; tests; README, AGENTS.md	cryptotavares/mm-cli-v2-extension 45c821d168a
Mobile tests/llm-workflow/	Same (write stays non-blocking)	cryptotavares/mm-cli-v2-mobile 4e2661f209f

Proof

Tested with a yalc build of this branch (0.12.0+f2caa574) in the Extension (e2e test build 13.56.0, headless) and Mobile (dev client 8.19.0, iOS sim + Android emulator), with the consumer commits above.

Check	Before (#59)	After
Files written under test-artifacts/llm-knowledge/ during a session	one step file per tool call	none (ext 116 → 116 files, mobile 260 → 260)
session.json per launch	llm-knowledge/<id>/session.json	test-artifacts/<id>/session.json (ext ×2, iOS, Android)
mm knowledge-last	step list	Error: unknown command 'knowledge-last', exit 1
mm launch --goal x	accepted	Warning: unknown launch flag '--goal', launch ok
Smoke (ext / iOS / Android)	pass	pass / pass / pass (Android list-testids 30 s timeout = known B11)
rg -i knowledge src	many	only the tests asserting the commands are gone
Coverage (stmts / branch / funcs / lines)	96.38 / 91.7 / 94.63 / 96.62	96.92 / 93.38 / 94.84 / 97.1

Transcripts

$ yarn mm knowledge-last
Error: unknown command 'knowledge-last'
exit=1

$ yarn mm launch --context e2e --state default --goal x
Warning: unknown launch flag '--goal'
... (launch ok, exit 0)

$ find test-artifacts -newer <marker> -type f      # Extension
test-artifacts/mm-mvl3vmad-zd909k/session.json
test-artifacts/mm-mvl3whgy-3sm2ye/session.json
test-artifacts/screenshots/p1-04-ext-home-1792768182733.png

$ cat test-artifacts/mm-mvl3vmad-zd909k/session.json
{ "schemaVersion": 1, "sessionId": "mm-mvl3vmad-zd909k", "createdAt": "...", "tags": [], "launch": { "stateMode": "default", ... } }

Unit tests

core:       Test Files 46 passed (46) · Tests 1397 passed (1397) · Type Errors no errors
            All files | 96.92 | 93.38 | 94.84 | 97.1
Extension:  yarn jest test/e2e/playwright/llm-workflow → 12 suites, 144 tests passed
Mobile:     yarn test:unit:llm-workflow → 18 suites, 231 tests passed

[seaona]

looks good, same question as previous comment: #59 (review)