docs(handoff): record Emma canonicalization baseline

This commit is contained in:
Chris Christiansen 2026-09-19 13:56:48 +00:00
parent c7c96efd12
commit 330e3ada1f

View File

@ -0,0 +1,193 @@
# Emma Canonicalization Handoff
## Scope
This handoff begins the controlled, phase-locked Emma canonicalization program.
The program uses one named phase per Gemini session, normally limited to
4560 minutes. Each phase must start from command-proven Git and runtime
state, stay inside its authorized scope, end with tests/diff/status, and stop.
## Verified Baseline
- **Repository:** `~/OSVauco`
- **Branch:** `feat/opax-domain-decouple`
- **Local HEAD:** `c7c96efd12bb3f3cdcfe9309e7683f0d01c16366`
- **Subject:** `fix(deploy): include MCP requirements in Cloud Build context`
- **Date:** `2026-09-19T12:39:31+00:00`
- **Local worktree status:** clean when this handoff was prepared
- **Remote status:** local `c7c96ef` was ahead of `origin/feat/opax-domain-decouple` before the handoff push
## Verified Public Health
- `https://opax.vauco.no/` returned HTTP 200 with valid TLS.
- `opax-mcp` authenticated health returned `status: ok`.
- The MCP health response reported the expected internal Ollama configuration.
- The public OPAX site and MCP service are distinct deployed services.
## Active Runtime
### OPAX MCP
- **Cloud Run service:** `opax-mcp`
- **Active revision:** `opax-mcp-00200-gcp`
- **Traffic:** 100%
- **Image:**
```text
us-central1-docker.pkg.dev/propane-will-491900-m5/osvauco-repo/opax-mcp@sha256:8687cd7489e789c17f6a56985191d6d924fc03d35d28c78bf94fea8b17b83091
```
### OPAX Web
- **Cloud Run service:** `opax-web`
- **Active revision:** `opax-web-00014-fnw`
- **Traffic:** 100%
- **Revision creation:** `2026-09-19T00:21:32.201043Z`
- **Image:**
```text
us-central1-docker.pkg.dev/propane-will-491900-m5/osvauco-repo/opax-web@sha256:27105067393c607d97a4c44fffe54bdf1da6ce7a6dd662e114f039fc04d69892
```
## Failed Acceptance Test
The live OPAX browser chat completed a basic response but failed the same-session
conversation-continuity acceptance test.
### Turn one
```text
User:
Husk dette testtokenet kun i denne samtalen:
OPAX-EMMA-1909.
Svar bare: registrert
Assistant:
registrert
```
### Turn two
```text
User:
Hva var testtokenet jeg ba deg huske?
Assistant:
Jeg har ikke godkjent VAUCO-kontekst i denne chatten ennå...
```
This is a failed same-session continuity acceptance test.
## Verified Local Source Path
The checked-out source contains the intended same-session history path:
```text
Browser conversation state
→ OPAX Web frontend request history
→ OPAX Web BFF
→ MCP JSON-RPC arguments.history
→ opax-mcp run_emma
→ _normalize_emma_history
→ CanonicalEmma.run
→ _ollama_chat
→ Ollama messages payload
```
Verified local implementation characteristics:
- `opax-mcp/server.py` `run_emma` passes normalized request history to `canonical_emma.run`.
- `_normalize_emma_history` retains only `user` and `assistant` roles.
- Blank messages are removed.
- The latest 40 historic messages are retained.
- `opax-mcp/emma_adapter.py` `CanonicalEmma.run` forwards history to its injected chat function.
- `opax-mcp/server.py` `_ollama_chat` places one canonical system prompt first, then history, then the current user prompt.
- Firestore/Morphic persistence is not part of the current same-session `run_emma` route.
## Unverified Runtime Facts
The following must remain explicitly marked as unverified until supported by
runtime payload evidence, deployed-source provenance, or focused tests:
- Whether active `opax-web` sends history in the real browser request.
- Whether active `opax-web` includes the local history implementation.
- Whether live Ollama receives the expected second-turn message sequence.
- Whether the active model uses valid history correctly after receiving it.
- Firestore/Morphic persistent-memory wiring.
- INCU ticket storage and execution wiring.
- Emma access to Gitea/Git context.
- Perplexity connector correctness.
## Architecture Boundaries
- OPAX Web is the authenticated UI and BFF.
- `opax-mcp` is the controlled model and tool gateway.
- Emma is the canonical runtime/agent.
- Same-session history, persistent memory, INCU ticketing, Git context, Git writes, and Perplexity connectivity are separate phases.
- The browser and model must not receive raw credentials, unrestricted terminal access, or direct infrastructure authority.
- Consequential actions require explicit human approval, resolved targets, structured arguments, audit evidence, and a visible result.
- Historic client messages may contain only `user` and `assistant` roles; historic `system` and `tool` roles must not reach the model payload.
## Phase Plan
1. **PHASE 0** — Runtime truth baseline.
2. **EMMA-SESSION-001A** — No-network session-history regression tests.
3. **EMMA-SESSION-001B** — Browser/BFF second-turn history-payload proof.
4. **EMMA-SESSION-001C** — Controlled release and live token acceptance.
5. **EMMA-PERSIST-001** — Authenticated, user-scoped Firestore session persistence.
6. **INCU-TICKET-001** — INCU ticket proposal/review workflow; no execution.
7. **EMMA-GIT-READ-001** — Structured read-only Git/Gitea context.
8. **EMMA-PATCH-001** — Patch proposal and allowlisted local validation.
9. **EMMA-GIT-WRITE-001** — Explicit approval-gated Git write lane.
10. **MCP-PERPLEXITY-001** — Separate Perplexity connector repair.
## Immediate Next Ticket
### EMMA-SESSION-001A
**Goal:** Add no-network regression coverage for the existing local history path.
Required test coverage:
1. `run_emma` normalizes and forwards valid historic `user` and `assistant` messages.
2. Historic `system`, `tool`, malformed, blank, and non-string-content entries are excluded.
3. `_ollama_chat` creates this exact final ordering:
```text
canonical system prompt
→ validated user/assistant history
→ current user prompt exactly once
```
**Out of scope:**
- Production code changes.
- Firestore or Morphic persistent memory.
- Git/Gitea access.
- INCU ticket execution.
- Commit, push, build, or deployment.
- Cloud Run, IAM, secrets, VPC, billing, or DNS changes.
**Completion condition:** Local no-network tests pass, diff is reviewed, and a human explicitly decides whether to commit.
## New Gemini Session Contract
Every Gemini session must:
1. Work on exactly one named phase.
2. Begin with raw output from:
- `pwd`
- `git branch --show-current`
- `git log -1`
- `git status --short`
3. Stop if the branch is not `feat/opax-domain-decouple`.
4. Stop if unexpected modifications are present.
5. Use exact source paths and line ranges.
6. Mark unsupported claims as `UNVERIFIED`.
7. Never invent Git SHAs, Cloud Run revisions, image digests, deployments, or test results.
8. End with actual test results, exact diff, Git status, confirmed facts, unverified facts, and a hard stop.
Commit, push, build, deployment, Firestore writes, Gitea writes, Cloud Run changes,
IAM changes, secret changes, VPC changes, and infrastructure actions require separate
explicit approval.