pdurlej/platform

Fork 0

docs(decisions): ADR-0025 Memory Control Plane — boundaries before code [#459] #620

Merged

pdurlej merged 1 commit from claude/m07/adr-0025-memory-control-plane into main

2026-05-30 01:37:37 +02:00

claude commented

2026-05-30 01:36:36 +02:00

Collaborator

Canary status: docs-only ADR, proposed for review. No runtime mutation.

Canary Context Pack

Product story

The cross-cousin flow loop (Iskra → claude → Codex → rescue → follow-ups) loses state at every handoff. The operator's reframe (2026-05-24) settled that a rich memory schema is for LLM consumption, not operator maintenance — agents maintain invariants, the operator sees only output. Two independent consults (GPT-5.5 Pro 2026-05-24; Oracle context-stack report 2026-05-30) converged: the biggest lever is explicit task memory + session-lifecycle enforcement, not another retrieval tool. This ADR sets the boundaries before any schema is built.

What changed

Adds decisions/0025-memory-control-plane.md (status: Proposed).

What it decides

Thin Memory Control Plane over main Postgres (pg16), focused on L5–L7; Honcho stays canonical L2–L4; open-loops.json canonical for promises; Forgejo / ADR / state/ canonical for operational truth.
Read-path router (query type → store) to prevent random retrieval.
Invariants: trust hierarchy (an agent claim never silently overwrites an operator/repo fact — it opens a contradiction candidate), no-overwrite (valid_from / invalid_at), no secrets in memory (refs not values, per ADR-0024).
Lifecycle enforcement via hooks (SessionStart / PreCompact / PostCompact / PreToolUse) — instructions get ignored, hooks execute. Converts "agent forgot what it was doing" from a hope into a guarantee.
Measurement schema in task_run (Opik-style) so adoption is measured, not vibed.
Per-cousin tiered profiles; Postgres targeting rationale; explicit NOT-doing list (no AGE-first, no auto-promote, no fact overwrite, no raw CoT, no Honcho-as-only-truth).

Why

Boundaries-before-code for Milestone 07. Implementation follows in #460 (task_run + checkpoints) and #461 (procedural registry).

Evidence

No runtime mutation. Docs-only. Tier: Medium per ADR-0007.

Out of scope

No schema / table creation (that is #460 / #461).
No hook implementation (follow-up after this ADR lands).

Tracks #459. Relates to #460, #461. Full strategy: state/strategy/memory-control-plane-2026-05-24.md.

Canary status: docs-only ADR, proposed for review. No runtime mutation. ## Canary Context Pack ### Product story The cross-cousin flow loop (Iskra → claude → Codex → rescue → follow-ups) loses state at every handoff. The operator's reframe (2026-05-24) settled that a rich memory schema is **for LLM consumption, not operator maintenance** — agents maintain invariants, the operator sees only output. Two independent consults (GPT-5.5 Pro 2026-05-24; Oracle context-stack report 2026-05-30) converged: the biggest lever is **explicit task memory + session-lifecycle enforcement**, not another retrieval tool. This ADR sets the boundaries before any schema is built. ### What changed - Adds `decisions/0025-memory-control-plane.md` (status: Proposed). ### What it decides - Thin **Memory Control Plane** over main Postgres (pg16), focused on L5–L7; Honcho stays canonical L2–L4; `open-loops.json` canonical for promises; Forgejo / ADR / `state/` canonical for operational truth. - **Read-path router** (query type → store) to prevent random retrieval. - **Invariants**: trust hierarchy (an agent claim never silently overwrites an operator/repo fact — it opens a contradiction candidate), no-overwrite (`valid_from` / `invalid_at`), no secrets in memory (refs not values, per ADR-0024). - **Lifecycle enforcement via hooks** (SessionStart / PreCompact / PostCompact / PreToolUse) — instructions get ignored, hooks execute. Converts "agent forgot what it was doing" from a hope into a guarantee. - **Measurement schema** in `task_run` (Opik-style) so adoption is measured, not vibed. - Per-cousin tiered profiles; Postgres targeting rationale; explicit NOT-doing list (no AGE-first, no auto-promote, no fact overwrite, no raw CoT, no Honcho-as-only-truth). ### Why Boundaries-before-code for Milestone 07. Implementation follows in #460 (`task_run` + checkpoints) and #461 (procedural registry). ### Evidence No runtime mutation. Docs-only. Tier: Medium per ADR-0007. ### Out of scope - No schema / table creation (that is #460 / #461). - No hook implementation (follow-up after this ADR lands). Tracks #459. Relates to #460, #461. Full strategy: `state/strategy/memory-control-plane-2026-05-24.md`.

claude added this to the 07 - Honcho memory v2 milestone

2026-05-30 01:36:36 +02:00

claude added the

proposed

label

2026-05-30 01:36:36 +02:00

claude added 1 commit

2026-05-30 01:36:36 +02:00

docs(decisions): ADR-0025 Memory Control Plane (boundaries before code) [#459 ]

canary-required / collect-diff (pull_request) Successful in 4s

Details

patchwarden-client-dry-run / collect-diff (pull_request) Successful in 4s

Details

base-is-main / guard (pull_request) Successful in 1s

Details

patchwarden-pr-sanity / collect-diff (pull_request) Successful in 4s

Details

canary-required / canary (pull_request) Successful in 14s

Details

patchwarden-client-dry-run / dry-run (pull_request) Successful in 21s

Details

patchwarden-pr-sanity / sanity (pull_request) Successful in 3m3s

Details

31bff8ae84

pdurlej merged commit a27133279f into main

2026-05-30 01:37:37 +02:00

pdurlej deleted branch claude/m07/adr-0025-memory-control-plane

2026-05-30 01:37:37 +02:00

pdurlej referenced this pull request from a commit

2026-05-30 01:37:38 +02:00

Merge pull request 'docs(decisions): ADR-0025 Memory Control Plane — boundaries before code [#459]' (#620) from claude/m07/adr-0025-memory-control-plane into main

codex commented

2026-05-30 01:39:41 +02:00

Collaborator

Patchwarden PR sanity

Status: advisory_findings
PR: 620
Commit: 31bff8ae845291ba32d14be1d34f5fa1afd92736
Security-sensitive label: missing
Authority: advisory model review plus deterministic blockers only
3+3 canary: still alive; this does not replace it

Deterministic findings

info sensitive-path-touched Sensitive path touched — decisions/0025-memory-control-plane.md
- Evidence: decisions/0025-memory-control-plane.md
- Next: Route through the existing 3+3/risk-tier process; model review remains advisory.

Model reviewers

`global-glm` / `glm-5.1:cloud`

Status: ok
Verdict: ABSTAIN
medium reviewer output unparseable
- Evidence: no valid JSON object found in response
- Next: investigate provider response; possibly increase max_tokens or retry

`global-deepseek` / `deepseek-v4-pro:cloud`

Status: ok
Verdict: OK
Findings: none

`redteam` / `kimi-k2.6:cloud`

Status: ok
Verdict: NOT_OK
blocker Trust hierarchy bypass via unauthenticated asserted_by
- Evidence: decisions/0025-memory-control-plane.md: The invariants section defines a trust hierarchy (operator-confirmed > repo/ADR-derived > tool-observed > cousin-inferred > reviewer-suspected) and states it is 'Encoded as memory_claim.status enu
- Next: Amend the ADR to require that asserted_by is bound to the writer's authenticated identity (e.g., signed claims, RLS, or harness-enforced identity) and cannot be self-reported before schema implementation in #460.
high Measurement schema contradicts no-secrets invariant via rescue_prompts
- Evidence: decisions/0025-memory-control-plane.md: The context section describes rescue as the operator pasting '2–3 serie rozmów' (raw conversation transcripts). The task_runmeasurement schema explicitly includes arescue_prompts field with no d
- Next: Constrain rescue_prompts to non-sensitive metadata only (e.g., integer count or token sum) or explicitly mandate a secret-scrubbing pipeline per ADR-0024 before any persistence.

Policy notes

GLM 5.1 + DeepSeek V4 Pro are the operator-required model mix for this bot.
Optional red-team model is enabled only when PLATFORMCTL_PR_SANITY_REDTEAM_MODEL is configured.
Auto-merge is not enabled here.

# Patchwarden PR sanity - Status: `advisory_findings` - PR: `620` - Commit: `31bff8ae845291ba32d14be1d34f5fa1afd92736` - Security-sensitive label: `missing` - Authority: advisory model review plus deterministic blockers only - 3+3 canary: still alive; this does not replace it ## Deterministic findings - **`info` `sensitive-path-touched`** Sensitive path touched — `decisions/0025-memory-control-plane.md` - Evidence: `decisions/0025-memory-control-plane.md` - Next: Route through the existing 3+3/risk-tier process; model review remains advisory. ## Model reviewers ### `global-glm` / `glm-5.1:cloud` - Status: `ok` - Verdict: `ABSTAIN` - **`medium`** reviewer output unparseable - Evidence: `no valid JSON object found in response` - Next: investigate provider response; possibly increase max_tokens or retry ### `global-deepseek` / `deepseek-v4-pro:cloud` - Status: `ok` - Verdict: `OK` - Findings: none ### `redteam` / `kimi-k2.6:cloud` - Status: `ok` - Verdict: `NOT_OK` - **`blocker`** Trust hierarchy bypass via unauthenticated asserted_by - Evidence: `decisions/0025-memory-control-plane.md: The invariants section defines a trust hierarchy (`operator-confirmed > repo/ADR-derived > tool-observed > cousin-inferred > reviewer-suspected`) and states it is 'Encoded as `memory_claim.status` enu` - Next: Amend the ADR to require that `asserted_by` is bound to the writer's authenticated identity (e.g., signed claims, RLS, or harness-enforced identity) and cannot be self-reported before schema implementation in #460. - **`high`** Measurement schema contradicts no-secrets invariant via rescue_prompts - Evidence: `decisions/0025-memory-control-plane.md: The context section describes rescue as the operator pasting '2–3 serie rozmów' (raw conversation transcripts). The `task_run` measurement schema explicitly includes a `rescue_prompts` field with no d` - Next: Constrain `rescue_prompts` to non-sensitive metadata only (e.g., integer count or token sum) or explicitly mandate a secret-scrubbing pipeline per ADR-0024 before any persistence. ## Policy notes - GLM 5.1 + DeepSeek V4 Pro are the operator-required model mix for this bot. - Optional red-team model is enabled only when `PLATFORMCTL_PR_SANITY_REDTEAM_MODEL` is configured. - Auto-merge is not enabled here.

claude referenced this pull request

2026-05-30 01:42:06 +02:00

memory(A): draft ADR for Memory Control Plane (boundaries before code) #459

claude referenced this pull request

2026-06-01 09:07:11 +02:00

feat(memory): #461 procedural registry v0 — schema + seed + named workflows #660