# QuietContext Product

audience: AI coding agents first.

status: ACTIVE — extraction slice DONE 2026-08-15 (contract lockdown 43 tests + inventory + carve landed, merged 9f06ca0, live daemon cut over and verified: healthz 200, tools/list 3967B parity both transports; insight/ proven NOT orphan and restored — inventory static tracing missed test-spawned subprocesses); daemon-hardening slice now ACTIVE (sonnet): savings ledger, burst hints (advise-at-2, max 3/window, no deny), idle-store eviction, orphaned tool_calls fix, execute-child CPU bounding.
task IDs: #176, #218–#225
source request: Simplify/bound QuietContext, clarify product purpose, compare architectures, write/review spec, plan implementation, then run through Factory.

## Outcome

QuietContext has a measured, minimal architecture that reduces conversation output/context without hidden reinstall overhead or unbounded indexing, then ships through Factory.

## Preserved progress

- #218 exploration marked complete; recover its output before repeating.
- #219 purpose clarification incomplete.
- #220–225 pending sequence: architecture comparison, spec, self-review, independent review, implementation plan, Factory run.
- #137 reportedly stopped `ensure-deps` reinstalling `better-sqlite3` on every hook; verify installed behavior before revisiting.
- Emergency lane forbids consuming current capacity until workstation flow is restored.

## Acceptance criteria

- Purpose and measurable success criteria approved.
- Architecture chosen from measured workloads.
- Resource/time/output bounds are explicit and enforced.
- No dependency install during normal hook calls.
- Factory implementation ships and is installed.

## Execution steps

1. Recover #218 findings.
2. Write purpose/success metrics here.
3. Compare smallest viable architectures with measurements.
4. Write one spec; self-review once; independent review once.
5. Implement through restored Factory.
6. Measure context/output/resource reduction.

## Current receipt

2026-08-15 owner ruling — architecture resolved: EXTRACTION, not in-place adjustment,
not greenfield. QuietContext is a context-mode fork still carrying dead code
(`registerLegacyTools`, `registerLegacyInsightTool`, unused paths in ~5,000-line
server.ts) and original inefficiencies; the July safe-core-cleanup deletion waves
never ran (`quietmode/docs/plans/2026-07-14-quietcontext-safe-core-cleanup.md` — its
contract-first wave plan remains the valid method). Sequence: the shared-HTTP-daemon
slice ships FIRST on the current core (transport-only over the six tool registrations,
see overdeck plan 2026-08-15-stateless-mcp-shared-daemon.md); the extraction slice then
carves only the live six-tool code paths into a clean core under the locked contract
tests — dead code drops by construction. The HTTP layer survives the swap unchanged.

## Next executable action

Recover #218 output and append its measured findings; never restart broad exploration.