pdurlej/platform

Fork 0

fix(honcho): route memory LLM defaults to qwen #699

Merged

pdurlej merged 3 commits from codex/690-honcho-qwen-fix into main

2026-06-03 00:13:07 +02:00

codex commented

2026-06-02 23:56:26 +02:00

Collaborator

Canary status: missing - fire canary 3+3 manually before merge

Summary

This PR takes over the repo-source slice from draft PR #690 and switches Honcho's default Ollama/OpenAI-compatible text model contract from gemma4:31b-cloud to qwen3.5.

It updates the source defaults, DR sandbox defaults, synthetic compatibility smoke, and active runbook/closeout docs. It does not restart production Honcho, does not edit RS2000 live override files, and does not write Infisical.

Refs #293, #357, #363, #690.

Canary Context Pack

Product story

Iskra/OpenClaw memory derivation must stop depending on a model that has been observed returning Markdown prose where Honcho needs JSON-compatible structured output. The repo source of truth should point future deploys and DR restores at the model already proven in the local classifier lane.

What changed

Changed Honcho LLM role defaults in compose/apps/compose.yaml from gemma4:31b-cloud to qwen3.5.
Aligned W3D local sandbox restore defaults with the same model.
Updated the Honcho Ollama compatibility smoke default/schema/prompt for Qwen while retaining the legacy script filename.
Updated tests for the new default model contract.
Updated active runbook, closeout plan, context MAP descriptions, and the incident packet with repo-source findings.

Why it changed

Draft PR #690 documents a live P1 symptom: Honcho derivation has been failing because the current Gemma route returns non-JSON prose. The source-of-truth repo needs to stop recreating that route before any runtime reconcile/restart happens.

Files touched

compose/apps/compose.yaml
scripts/dr/w3d-local-sandbox-drill.sh
scripts/honcho/ollama-gemma-compat-smoke.py
control-plane/platformctl/tests/test_honcho_ollama_contract.py
tests/test_honcho_log_privacy.py
runbooks/honcho-ollama-gemma-switch.md
state/cutover/honcho-closeout-plan.md
contexts/persona-bridge/MAP.md
contexts/observability/MAP.md
docs/incidents/2026-06-02-honcho-qwen-fix.md

Relevant context

Draft PR #690 incident packet.
Platform issue #357 for embedding-space migration boundaries.
Platform issue #363 / Memory OS spine context.
Existing Honcho Ollama provider switch runbook and closeout plan.

Runtime evidence

Repo-only evidence in this PR:

PYTHONPATH=control-plane python3 -m pytest control-plane/platformctl/tests/test_honcho_ollama_contract.py tests/test_honcho_log_privacy.py passed.
git diff --check passed.
bash -n scripts/dr/w3d-local-sandbox-drill.sh passed.
PYTHONPATH=control-plane python3 -m platformctl.cli validate all --json passed with exitCode: 0.

No production restart, no live Infisical write, and no live Honcho derivation proof was performed in this PR.

Known constraints

The live acceptance criteria from #690 still require a gated runtime step: reconcile Infisical and RS2000 live overrides, restart only home-platform-honcho-deriver-1 and home-platform-honcho-api-1, then prove valid JSON derivation and backlog progress.

Explicit out-of-scope

No production Honcho restart.
No live override edit on RS2000.
No Infisical write.
No live Postgres query or migration.
No embedding-space migration.
No attempt to close #357.

Requested decision

Approve this as the source-of-truth repo fix needed before the runtime cutover/reconcile step.

Merge blockers

If reviewers find another active Honcho source-of-truth still defaulting to gemma4:31b-cloud.
If the Qwen tag is disputed; current local evidence points to exact model tag qwen3.5 from the available classifier integration checkout.
If canary finds this should be split from the DR/runbook alignment.

Spec sources read

docs/incidents/2026-06-02-honcho-qwen-fix.md - incident packet and acceptance constraints.
compose/apps/compose.yaml - Honcho LLM defaults source of truth.
scripts/dr/w3d-local-sandbox-drill.sh - DR/sandbox restore default contract.
scripts/honcho/ollama-gemma-compat-smoke.py - synthetic provider compatibility smoke.
control-plane/platformctl/tests/test_honcho_ollama_contract.py - Honcho model/default tests.
tests/test_honcho_log_privacy.py - model metadata in log privacy fixture.
runbooks/honcho-ollama-gemma-switch.md - operator runbook.
state/cutover/honcho-closeout-plan.md - active closeout plan.
contexts/persona-bridge/MAP.md and contexts/observability/MAP.md - context navigation labels.
/Users/pd/Developer/books-for-iskra/iskra-openclaw-integration/scripts/iskra-fastmail-dmz-classifier.py - exact existing Qwen model tag evidence.

Canary status: missing - fire canary 3+3 manually before merge ## Summary This PR takes over the repo-source slice from draft PR #690 and switches Honcho's default Ollama/OpenAI-compatible text model contract from `gemma4:31b-cloud` to `qwen3.5`. It updates the source defaults, DR sandbox defaults, synthetic compatibility smoke, and active runbook/closeout docs. It does not restart production Honcho, does not edit RS2000 live override files, and does not write Infisical. Refs #293, #357, #363, #690. ## Canary Context Pack ### Product story Iskra/OpenClaw memory derivation must stop depending on a model that has been observed returning Markdown prose where Honcho needs JSON-compatible structured output. The repo source of truth should point future deploys and DR restores at the model already proven in the local classifier lane. ### What changed - Changed Honcho LLM role defaults in `compose/apps/compose.yaml` from `gemma4:31b-cloud` to `qwen3.5`. - Aligned W3D local sandbox restore defaults with the same model. - Updated the Honcho Ollama compatibility smoke default/schema/prompt for Qwen while retaining the legacy script filename. - Updated tests for the new default model contract. - Updated active runbook, closeout plan, context MAP descriptions, and the incident packet with repo-source findings. ### Why it changed Draft PR #690 documents a live P1 symptom: Honcho derivation has been failing because the current Gemma route returns non-JSON prose. The source-of-truth repo needs to stop recreating that route before any runtime reconcile/restart happens. ### Files touched - `compose/apps/compose.yaml` - `scripts/dr/w3d-local-sandbox-drill.sh` - `scripts/honcho/ollama-gemma-compat-smoke.py` - `control-plane/platformctl/tests/test_honcho_ollama_contract.py` - `tests/test_honcho_log_privacy.py` - `runbooks/honcho-ollama-gemma-switch.md` - `state/cutover/honcho-closeout-plan.md` - `contexts/persona-bridge/MAP.md` - `contexts/observability/MAP.md` - `docs/incidents/2026-06-02-honcho-qwen-fix.md` ### Relevant context - Draft PR #690 incident packet. - Platform issue #357 for embedding-space migration boundaries. - Platform issue #363 / Memory OS spine context. - Existing Honcho Ollama provider switch runbook and closeout plan. ### Runtime evidence Repo-only evidence in this PR: - `PYTHONPATH=control-plane python3 -m pytest control-plane/platformctl/tests/test_honcho_ollama_contract.py tests/test_honcho_log_privacy.py` passed. - `git diff --check` passed. - `bash -n scripts/dr/w3d-local-sandbox-drill.sh` passed. - `PYTHONPATH=control-plane python3 -m platformctl.cli validate all --json` passed with `exitCode: 0`. No production restart, no live Infisical write, and no live Honcho derivation proof was performed in this PR. ### Known constraints The live acceptance criteria from #690 still require a gated runtime step: reconcile Infisical and RS2000 live overrides, restart only `home-platform-honcho-deriver-1` and `home-platform-honcho-api-1`, then prove valid JSON derivation and backlog progress. ### Explicit out-of-scope - No production Honcho restart. - No live override edit on RS2000. - No Infisical write. - No live Postgres query or migration. - No embedding-space migration. - No attempt to close #357. ### Requested decision Approve this as the source-of-truth repo fix needed before the runtime cutover/reconcile step. ### Merge blockers - If reviewers find another active Honcho source-of-truth still defaulting to `gemma4:31b-cloud`. - If the Qwen tag is disputed; current local evidence points to exact model tag `qwen3.5` from the available classifier integration checkout. - If canary finds this should be split from the DR/runbook alignment. ## Spec sources read - `docs/incidents/2026-06-02-honcho-qwen-fix.md` - incident packet and acceptance constraints. - `compose/apps/compose.yaml` - Honcho LLM defaults source of truth. - `scripts/dr/w3d-local-sandbox-drill.sh` - DR/sandbox restore default contract. - `scripts/honcho/ollama-gemma-compat-smoke.py` - synthetic provider compatibility smoke. - `control-plane/platformctl/tests/test_honcho_ollama_contract.py` - Honcho model/default tests. - `tests/test_honcho_log_privacy.py` - model metadata in log privacy fixture. - `runbooks/honcho-ollama-gemma-switch.md` - operator runbook. - `state/cutover/honcho-closeout-plan.md` - active closeout plan. - `contexts/persona-bridge/MAP.md` and `contexts/observability/MAP.md` - context navigation labels. - `/Users/pd/Developer/books-for-iskra/iskra-openclaw-integration/scripts/iskra-fastmail-dmz-classifier.py` - exact existing Qwen model tag evidence.

codex added 2 commits

2026-06-02 23:56:26 +02:00

WIP: fix(honcho): swap gemma4→qwen3.5 — memory derivation broken (P1) 5dbac0106e

fix(honcho): route memory LLM defaults to qwen

canary-required / collect-diff (pull_request) Successful in 5s

Details

platformctl plan / auto-apply scope (pull_request) Successful in 20s

Details

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

Details

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

Details

patchwarden-pr-sanity / sanity (pull_request) Has been cancelled

Details

base-is-main / guard (pull_request) Waiting to run

Details

patchwarden-pr-sanity / collect-diff (pull_request) Waiting to run

Details

pyfallow / Pyfallow gate (control-plane) (pull_request) Successful in 19s

Details

python-ci / Python 3.11 (pull_request) Successful in 41s

Details

python-ci / Python 3.12 (pull_request) Successful in 46s

Details

python-ci / Python 3.13 (pull_request) Successful in 46s

Details

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

Details

c094cd8175

codex added the

labels

2026-06-02 23:56:58 +02:00

codex referenced this pull request

2026-06-02 23:56:59 +02:00

WIP: fix(honcho): swap gemma4→qwen3.5 — memory derivation broken (P1) #690

codex referenced this pull request

2026-06-02 23:59:13 +02:00

ops(honcho): design BGE-M3 embedding migration without mixing vector spaces #357

codex added 1 commit

2026-06-03 00:03:23 +02:00

chore(ci): retrigger honcho qwen checks

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

Details

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

Details

pyfallow / Pyfallow gate (control-plane) (pull_request) Successful in 18s

Details

python-ci / Python 3.11 (pull_request) Successful in 41s

Details

python-ci / Python 3.12 (pull_request) Successful in 42s

Details

patchwarden-client-dry-run / collect-diff (pull_request) Successful in 3s

Details

patchwarden-pr-sanity / collect-diff (pull_request) Successful in 3s

Details

platformctl plan / auto-apply scope (pull_request) Successful in 19s

Details

python-ci / Python 3.13 (pull_request) Successful in 41s

Details

canary-required / canary (pull_request) Successful in 13s

Details

patchwarden-client-dry-run / dry-run (pull_request) Successful in 17s

Details

patchwarden-pr-sanity / sanity (pull_request) Successful in 2m51s

Details

deb5081804

codex commented

2026-06-03 00:09:42 +02:00

Author

Collaborator

Patchwarden PR sanity

Status: advisory_findings
PR: 699
Commit: deb5081804ccee2c3dbbbc04379d37ae4eba3fd2
Security-sensitive label: missing
Authority: advisory model review plus deterministic blockers only
3+3 canary: still alive; this does not replace it

Deterministic findings

No deterministic findings.

Model reviewers

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

Status: ok
Verdict: OK
medium Legacy filenames may confuse future operators searching for Qwen references
- Evidence: scripts/honcho/ollama-gemma-compat-smoke.py retains 'gemma' in filename while testing qwen3.5; runbooks/honcho-ollama-gemma-switch.md retains 'gemma' in filename while documenting Qwen switch. Diff shows schema changed to 'honcho_ollama_qwe
- Next: Consider adding a prominent comment at the top of ollama-gemma-compat-smoke.py explaining the legacy filename retention, or create a symlink ollama-qwen-compat-smoke.py -> ollama-gemma-compat-smoke.py for discoverability. The runbook filename is lower risk since it's documentation, but a similar not

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

Status: ok
Verdict: OK
low Model tag qwen3.5 not verified against Ollama Cloud catalog
- Evidence: compose/apps/compose.yaml: all Honcho model defaults changed to qwen3.5; incident doc references local classifier checkout but no live Ollama Cloud tag confirmation.
- Next: Before runtime cutover, confirm that qwen3.5 is the exact tag available on Ollama Cloud and that it supports the required structured output for all Honcho roles.
low Smoke script filename still references Gemma
- Evidence: scripts/honcho/ollama-gemma-compat-smoke.py: filename unchanged while content now defaults to Qwen; runbooks still reference this filename.
- Next: Plan a follow-up rename of the script to ollama-qwen-compat-smoke.py after verifying no remote automation hardcodes the old name.

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

Status: ok
Verdict: NOT_OK
high Compatibility smoke falls back to plain prompt, masking JSON structured-output failures
- Evidence: scripts/honcho/ollama-gemma-compat-smoke.py chat_test only validates a plain string echo (HONCHO_QWEN_OK) and does not enforce JSON schema; control-plane/platformctl/tests/test_honcho_ollama_contract.py test_ollama_smoke_json_test_falls_bac
- Next: Remove the plain-prompt fallback from smoke.json_test (or add a separate structured-output smoke that hard-fails on non-JSON) and update the runbook acceptance criteria to require 100% valid JSON before DR/closeout.

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: `699` - Commit: `deb5081804ccee2c3dbbbc04379d37ae4eba3fd2` - Security-sensitive label: `missing` - Authority: advisory model review plus deterministic blockers only - 3+3 canary: still alive; this does not replace it ## Deterministic findings No deterministic findings. ## Model reviewers ### `global-glm` / `glm-5.1:cloud` - Status: `ok` - Verdict: `OK` - **`medium`** Legacy filenames may confuse future operators searching for Qwen references - Evidence: `scripts/honcho/ollama-gemma-compat-smoke.py retains 'gemma' in filename while testing qwen3.5; runbooks/honcho-ollama-gemma-switch.md retains 'gemma' in filename while documenting Qwen switch. Diff shows schema changed to 'honcho_ollama_qwe` - Next: Consider adding a prominent comment at the top of ollama-gemma-compat-smoke.py explaining the legacy filename retention, or create a symlink ollama-qwen-compat-smoke.py -> ollama-gemma-compat-smoke.py for discoverability. The runbook filename is lower risk since it's documentation, but a similar not ### `global-deepseek` / `deepseek-v4-pro:cloud` - Status: `ok` - Verdict: `OK` - **`low`** Model tag `qwen3.5` not verified against Ollama Cloud catalog - Evidence: `compose/apps/compose.yaml: all Honcho model defaults changed to `qwen3.5`; incident doc references local classifier checkout but no live Ollama Cloud tag confirmation.` - Next: Before runtime cutover, confirm that `qwen3.5` is the exact tag available on Ollama Cloud and that it supports the required structured output for all Honcho roles. - **`low`** Smoke script filename still references Gemma - Evidence: `scripts/honcho/ollama-gemma-compat-smoke.py: filename unchanged while content now defaults to Qwen; runbooks still reference this filename.` - Next: Plan a follow-up rename of the script to `ollama-qwen-compat-smoke.py` after verifying no remote automation hardcodes the old name. ### `redteam` / `kimi-k2.6:cloud` - Status: `ok` - Verdict: `NOT_OK` - **`high`** Compatibility smoke falls back to plain prompt, masking JSON structured-output failures - Evidence: `scripts/honcho/ollama-gemma-compat-smoke.py chat_test only validates a plain string echo (HONCHO_QWEN_OK) and does not enforce JSON schema; control-plane/platformctl/tests/test_honcho_ollama_contract.py test_ollama_smoke_json_test_falls_bac` - Next: Remove the plain-prompt fallback from smoke.json_test (or add a separate structured-output smoke that hard-fails on non-JSON) and update the runbook acceptance criteria to require 100% valid JSON before DR/closeout. ## 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.

pdurlej approved these changes

2026-06-03 00:12:39 +02:00

pdurlej left a comment

Operator-delegated approval via temporary admin lane: PR #699 is repo-source only, all required checks are green, no runtime restart, no Infisical write, no live override edit.