[cryptotavares]

Summary: Remove priorKnowledge from describe_screen and its public observation types. Screen discovery no longer reads historical sessions to generate suggestions on each call.

Stack context: mm CLI v2, Phase 1, P1-03. Base: cryptotavares/mmcli-p1-02-device-mcp-0-5; review #58 first. #60 removes the remaining knowledge store.

Problem

Each describe_screen call gathers the current screen, then asks the knowledge store for historical context. That lookup reads past step files and appends related sessions, similar steps, and suggested actions to the result.

This couples current-screen discovery to history size on disk. A separate change must isolate that lookup's cost before the remaining knowledge store is removed.

Solution

Remove the history-generation call at the shared discovery-tool boundary. Browser, iOS, and Android still collect current-screen data through their platform driver.

flowchart TD
    A[describe_screen] --> B[Collect state, test IDs, and accessibility tree]
    B --> C[Refresh accessibility refs and resolve active tab]
    C --> D[Capture screenshot when requested]
    D --> E[Return current-screen data without priorKnowledge]

    classDef new fill:#fff3cd,stroke:#d39e00
    class E new

Loading

Yellow boxes are new.

Remove the same field from the result and step-observation types, observation diffing, and prior-knowledge redaction. createDefaultObservation now returns its three-field observation directly.

What changes

Condition or boundary	Before	After
describe_screen history access	Calls generatePriorKnowledge after screen collection	Returns without that call
Screen result and step-observation contracts	Optional priorKnowledge field	Field removed; current-screen data remains
createDefaultObservation	Optional fourth argument	Three arguments: state, test IDs, accessibility nodes
Observation diffing and step redaction	Carry or scrub prior knowledge	Remove prior-knowledge-specific handling
Remaining knowledge-store features	Step recording, explicit knowledge tools, and history-generation API	Retained for the next PR

Decisions for reviewers

Decision	Reason
Split P1-03 from the full store removal	Isolates the history lookup and payload cost. The protected proof below records the original measurements.
Remove the field from both public types	Keeps discovery results and recorded observations aligned with the producer. This is a breaking contract change.
Retain the knowledge store in this PR	Keeps step recording and explicit knowledge queries available until #60.

Review guide

Read the files in this order:

#	File	What to check
1	src/tools/discovery-tools.ts	describeScreenTool returns current-screen data without consulting the knowledge store.
2	src/tools/types/tool-outputs.ts, src/tools/types/step-record.ts	The two public contracts no longer expose priorKnowledge.
3	src/knowledge-store/knowledge-store.ts	createDefaultObservation accepts three arguments; the remaining store stays available.
4	src/server/observation-compaction.ts, src/knowledge-store/utils/redaction.ts	Remove only prior-knowledge propagation and redaction.
5	Changed tests and src/tools/test-utils/mock-factories.ts	Discovery, driver delegation, recording, redaction, and compaction match the reduced contract.
6	README.md, SKILL.md	Remove historical-suggestion claims from the discovery instructions.

Testing

• Fresh runtime check at remote head 6a20fa8c3cc3af828aedd6612db36590667c4952: 7 files, 369 tests pass:

  CI=true yarn vitest run src/tools/discovery-tools.test.ts src/tools/driver-delegation.test.ts src/knowledge-store/knowledge-store.test.ts src/knowledge-store/utils/redaction.test.ts src/server/observation-compaction.test.ts src/server/create-server.test.ts src/tools/test-utils/mock-factories.test.ts --coverage.enabled=false --typecheck.enabled=false --maxWorkers=2 --minWorkers=1

• yarn tsc --noEmit -p tsconfig.build.json passes. ESLint passes for all 14 changed TypeScript files.

• yarn prettier README.md SKILL.md --check passes.

• yarn tsc --noEmit -p tsconfig.test.json fails with inherited TS1309 at src/platform/mobile-platform-driver.hermes.test.ts:25. That file is identical to the immediate parent.

• The Proof and Unit tests sections below retain the original historical evidence.

Breaking changes

• describe_screen results (mm describe-screen) no longer contain priorKnowledge.
• The DescribeScreenResult and StepRecordObservation types no longer have priorKnowledge.
• createDefaultObservation(state, testIds, nodes) no longer takes a 4th priorKnowledge argument.

Consumer impact

Consumer	Change needed	Done on branch
Extension test/e2e/playwright/llm-workflow/	None (rg priorKnowledge: 0 hits)	n/a
Mobile tests/llm-workflow/	None (rg priorKnowledge: 0 hits)	n/a

Proof

Tested with yalc builds of the parent (c87dbc3) and of this branch in the Extension (e2e test build 13.56.0, headless) and Mobile (dev client 8.19.0, iOS sim + Android emulator), on wallet home after unlock.

Check	Before	After
priorKnowledge in describe-screen (ext / iOS / Android)	present (1/1/1)	absent (0/0/0)
Size of the removed block (pretty-printed, ext / iOS / Android)	3881 / 5060 / 4856 B	0
describe-screen bytes p50, Extension	14019	9816
describe-screen toolMs p50, Extension (5 runs)	712	614
describe-screen toolMs p50, Android (3 runs)	6916	4034
generatePriorKnowledge alone, Mobile store (18 sessions, 31 MB)	p50 294 ms per call	0
generatePriorKnowledge alone, Extension store (9 sessions, 1 MB)	p50 30 ms per call	0

The iOS device bench was too noisy on this host (load average 11–19) to show the ~0.3 s gain: observationMs, which this change does not touch, moved by the same amount as toolMs. The isolated timing of the removed call is the reliable number. Its cost grows with the history on disk.

Transcripts

describe-screen bench p50 (bytes / wall / toolMs / observationMs / totalMs)
ext      before 14019 / 2177 / 712 / 634 / 1348     after 9816 / 2182 / 614 / 642 / 1242
android  before 1010607 / 10957 / 6916 / 2928 / 9849   after 1014571 / 8457 / 4034 / 2845 / 7355
(Android a11y tree on wallet home varies by ~7 kB between runs, which hides the 4.9 kB removal)

$ grep -c priorKnowledge {ext,ios,android}/describe-screen.txt
0 0 0
Top-level keys after: state, activeTab (ext), testIds, a11y, screenshot, meta

Unit tests

 Test Files  49 passed (49)
      Tests  1636 passed (1636)
Type Errors  no errors
All files          |   96.38 |     91.7 |   94.63 |   96.62 |

[seaona]

hey there 👋 the removal code looks good.
Just a question to further develop, are there any plans to replace this with a different (more performant) approach in the future?
So from what I undestand, now we are transitioning towards a stateless tool. However, we could move towards agent memory or teh saved recipies/scripts we talked about before.
Is the path forward documented somewhere / decided?

[cryptotavares]

hey there 👋 the removal code looks good. Just a question to further develop, are there any plans to replace this with a different (more performant) approach in the future? So from what I undestand, now we are transitioning towards a stateless tool. However, we could move towards agent memory or teh saved recipies/scripts we talked about before. Is the path forward documented somewhere / decided?

I believe that a knowledge store or memory should not be tied to the tool itself. The tool should be a stateless tool. Agents knowledge/memory, should be tackled separately. In MetaMask scope, maybe up to date feature maps is all we need