- new context minimax-m31-flash/v1 (model pin; no thinking-budget init, Flash
ignores it); four composites switch from minimax-m3-max/v1 (kept for rollback)
- models + registry for minimax-m31-flash-{max,medium}; removes the dangling
model-registry/minimax.yaml (endpoint "minimax" never existed)
- tiers, harness-roles, minimax-* task templates and the *-review-loop@1
workflows use the Flash slugs (MiniMax silently maps unknown names to M3, so
the provider-id translation lands in agent-runtimes)
init.sh now archives and resets an existing task/<uuid> branch to the current
base when AGENT_BRANCH_RESET_ON_START=true. Default off: all nodes of a
workflow share the branch, so only the workflow's first node may set it.
Claude-Session: https://claude.ai/code/session_01N218bERyNeLhbLmDRDvbWF
Hatchmate templates resolving to claude-sonnet-5, claude-haiku-4-5 or
airouter-deepseekv4flash mounted only gitea-ssh/v1 and could not clone
hatchmate/hatchmate-agents. Mirror the existing -hatchmate composites:
non-hatchmate base layers plus gitea-ssh-hatchmate/v1 after gitea-ssh/v1 (HC-26).
Claude-Session: https://claude.ai/code/session_01N218bERyNeLhbLmDRDvbWF
nocp: shared-variables/v1 + repo-readonly/v1 (no CP credential).
write: shared-variables/v1 + cp-write-token/v1 + repo-readonly/v1.
repo-readonly/v1 is last in both (WF-COMP-34); neither layers
cp-service-token/v1 or gitea-ssh/v1.
HOLD: do not merge until the CP release (AU-55/55a/56/57, WF-COMP-31b engine/
dispatch/compiler composite-set changes) and the dispatcher image
(cp-write-token delivery, AU-55b) are deployed. The dispatcher advertises every
composite found in the CRS-synced harnesses/ tree (dispatcher/poller.py
_collect_supported_harnesses), so merging this to origin/main IS the
advertisement event.
Claude-Session: https://claude.ai/code/session_01N218bERyNeLhbLmDRDvbWF
Env-only layer exporting CP_WRITE_TOKEN_FILE=/run/agent/cp-write-token/token.
The per-step wfcw_ token is delivered by the dispatcher into the script
container only (AU-55b), so the layer has no secrets_required, init script or
files. Not referenced by any composite yet; safe to merge at any time.
Claude-Session: https://claude.ai/code/session_01N218bERyNeLhbLmDRDvbWF
repo-readonly/v1 now declares write_targets: [], an ssh_hosts alias and its
own secrets_required (account platform-readonly.space:platform, a read-only
Gitea deploy key), and stages the key before cloning. workflow-script/v1 no
longer layers gitea-ssh/v1, so script nodes cannot push. Invariants updated:
credential layers in the composite must declare write_targets: [].
Bug 45a48685-23e9-4605-af1f-d93306108cef
Claude-Session: https://claude.ai/code/session_01N218bERyNeLhbLmDRDvbWF
init.sh invoked 'python3 -m entrypoint.agent_repo_base' from /workspace/project, so a project
repo's own entrypoint/ shadowed the image's and the selector crashed. Invoke the image copy by
path. scope-decompose-sonnet@2 gains AGENT_UPSTREAM_REPO_URL/BRANCH so fresh spec branches seed
from upstream main instead of the stale fork main.
Claude-Session: https://claude.ai/code/session_01N218bERyNeLhbLmDRDvbWF
M2-CRS W4 unblock: dispatched agents could not clone hatchmate/hatchmate-agents
because the shared gitea-ssh/v1 context mounts the ai_enablement deploy key,
which is scoped to skynet/* only. Same bug would hit every project whose
working repo lives outside the skynet org.
This implements the per-org pattern the framework already uses for accelerators
(gitea-ssh-accelerators/v1) and homelab (gitea-ssh-homelab/v1):
- harnesses/contexts/gitea-ssh-hatchmate/v1/{init.sh,harness.yaml} — NEW
forge-credential context mirroring accelerators, with account_ref
'hatchmate.space:platform' and write_targets covering hatchmate/*. The
init.sh stages the ESO-mounted private key to ~/.ssh/gitea-oreillyit-nz-hatchmate
(matching the ssh alias declared in harness.yaml). URL host pattern
'gitea.oreillyit.nz-hatchmate' (per-org alias) drives SSH config block
selection — already auto-written by harness_init.py from each layer's
ssh_hosts: declarations.
- harnesses/composites/code-*-tdd-repo-hatchmate/v1/ — NEW composites for
fable/minimax/airouter/opus that compose BOTH gitea-ssh/v1 (for reference
clones, skynet org) AND gitea-ssh-hatchmate/v1 (for working-repo clone +
push, hatchmate org). Composites are unchanged for non-hatchmate projects.
- tests/invariants/test_forge_credentials_declare_write_targets.py — the
existing xfail(strict=True) marker was wrong: the 4 existing layers all
declare write_targets today, so the test was XPASSing. Removed the
xfail marker so HC-11 is now a real positive gate (5 layers instead
of 4). This was a pre-existing test bug surfaced by the 5th layer.
W4 next step: dispatcher's account_ref resolution needs the hatchmate CP
account + grant to be created (via the new scripts/provision-hatchmate-fork-key.sh
in the hatchmate app repo).
uv is now pre-baked into the base image
(agent-runtimes/images/base/Dockerfile) and is faster than pip with
better retry on transient network failures. Closes the last runtime
hang vector for project-deps install — apt install was removed 2026-09-06
in 0c57972f-pre-bake work but the pip install -e line in init.sh
remained. uv pip install is the equivalent operation with the new
pre-baked tooling.
Refs F-WE-PDIV/RCA-2026-09-04-zombies.
verdict-tx (kind: script) never worked: script nodes are harness-exempt
(WF-COMP-6/MN-38) and run in their own empty container, so it could
never credentially read /workspace/verdict from the agent's node.
Option 3: give the agent node's own composite (planning-sonnet-repo) a
cp-service-token/v1 layer, and have the agent PATCH its own item's
flow_state (verified|rejected) to the CP directly as its last step,
before task_complete. Removes verdict-tx entirely.
See agent-runtimes work item 88ecd23e and spec/typed-workflows.md
WF-COMP-25 (agent-runtimes commit 8972d266) for the finalize.py
precedence fix this depends on (ambiguous pipe-separated output.state
must never be auto-PATCHed over the agent's own value).
Claude-Session: https://claude.ai/code/session_01U3biMxRhFGkwtf4jVTNjcj
Minimal harness for kind: customer_script nodes (agent-runtimes
spec/tenant-namespaces.md SCR-1..SCR-6). No LLM turn - runs via the
"shell" runner, agent-repo/v1 clone access only. Per-script credential
needs are deliberately not baked into this base (SCR-6: narrow,
per-script review) - compose a customer-script-<need>/v1 variant
layering whichever specific credential context a given script actually
needs, rather than widening this shared base.
Without this commit, kind: customer_script was not dispatchable from
any machine other than the one it was authored on - the harness lookup
against a fresh CRS sync would fail.
Claude-Session: https://claude.ai/code/session_012arAwUuBQrpnMzJ36mEVBk
Cross-node artifact passing within a workflow manifest — nodes each run
in their own ephemeral container/pod with no shared filesystem, so an
agent node writing e.g. a verdict to /workspace/verdict is invisible to
a downstream kind: script node in the same manifest (found live
2026-09-06, bug-verify-sonnet's verdict-tx). The git task branch is the
only channel every node in a manifest actually shares: write to
scratchpad/, commit (picked up by finalize.sh's existing git add -A),
next node clones the same branch and reads it back.
scratchpad/ is scratch, never deliverable content, and must never
survive a merge into main — a CI gate in the target repo
(agent-runtimes .gitea/workflows/build-images.yaml,
no-scratchpad-in-main) is the hard backstop.
Adds the CRS-side deliverables for agent-runtimes work item f59f0132
(ai-resolve bug-verification stage, M42 concept 7666d278). Mirrors
planning-opus-repo's layer stack with the sonnet model applied at the
node level (model: sonnet), per the established convention that the
composite name is descriptive only — the workflow node's own model:/
tier: field is what actually selects the model.
The template's terminal node runs the (separately-tracked, sibling
task ad975ac6) state-tx-verdict script — this template will not be
runnable end-to-end until that script lands, but the declarative
config itself is independently valid: schema-validates cleanly and
both eligibility scenarios from the task spec pass (verify-pending +
ai-resolve -> ok=True; missing flow_state -> state_mismatch).
The ESO-mounted secret is root-owned mode 0400 (correct — init.sh runs as
root before the gosu drop to the agent user). The original version of
this context pointed CP_SERVICE_TOKEN_FILE directly at that raw mount,
which the agent process (uid 1000) can never read.
Confirmed live 2026-09-03: a real scope-decompose-sonnet@1 dispatch
against concept 7666d278-ae58-4f12-990d-c4959a3e19a9 hit exactly this —
the agent correctly diagnosed 'cp_cli invocation can't proceed... token
is root-owned mode 0400 so it's not readable by the agent user', wrote
its decompose plan, but could never actually call cp-cli to create the
child tasks. The trigger engine's new side-effect verification (finalize.py,
bug 9dffc5b8) correctly caught this and refused to advance flow_state —
so this was a visible, retriable failure rather than another silent
false-positive.
Fix: init.sh restages the secret to /run/agent/cp-service-token/token,
mode 0600, owned by the agent user — mirrors
anthropic-cloud-paul-oauth/v1/init.sh's existing pattern.
CP_SERVICE_TOKEN_FILE now points at the restaged copy.
New context mounts the cp-decompose-service-token account (system
provider type, account_id 1d963673-6ac9-4f85-875a-2ce5323e76ad, owner
(cp, cp)) via ESO and exports CP_SERVICE_TOKEN_FILE pointing at the
mounted secret file. Wired into spec-writing-sonnet-repo (backs
scope-decompose-sonnet@1, which calls back POST
/v1/work-items/{id}/decompose — bug b3a96acc / fa41caac / M42 stall).
The mounted value must match the CP's CP_INTERNAL_BEARER_TOKEN
(agent-runtimes deploy repo, overlays/{production,staging}/app-secret.sops.yaml).
No init.sh needed — the mount path is static at harness-authoring time
and entrypoint/actions/decompose_work_items.py (AU-51b, agent-runtimes
commit 0fa8b760) reads CP_SERVICE_TOKEN_FILE directly, so no
transformation/staging step is required.
'requires: [claude-code]' failed composition ('Invalid harness reference
format'). Both contexts need no hard dependency — layer ordering in the
composites already guarantees the auth layer precedes them. Validated
locally with entrypoint.harness compose_harness for all 6 affected
composites this time.
Claude-Session: https://claude.ai/code/session_019tJk7P8tZzJ24PvtgoGhLN
The validator pattern-matches *TOKEN* as credential-shaped and rejected
both contexts, failing every task on the new harnesses. The env var name
is fixed by Claude Code, so deliver the thinking budget via the
settings.json env map in init.sh instead (jq-merge preserves apiKeyHelper
from the minimax/oauth layer).
Claude-Session: https://claude.ai/code/session_019tJk7P8tZzJ24PvtgoGhLN
Integration runs were burning the personal Anthropic 5-hour window
(429 storm + window drain, 2026-09-02). New env-only context
minimax-m3-max/v1 pins ANTHROPIC_MODEL=MiniMax-M3[1m], maxes the
thinking budget, and sets the 1M auto-compact window; layered on
minimax/v1 for base URL + apiKeyHelper auth.
Claude-Session: https://claude.ai/code/session_019tJk7P8tZzJ24PvtgoGhLN
The 237KB library was mounted at /workspace/best-practices/ in every
consuming ConfigMap but nothing in the prompt mentioned it (directory
targets never enter the context manifest). New ~90-word CLAUDE.md file
target enters the system prompt: points at INDEX.md, instructs reading
only task-relevant topics.
Claude-Session: https://claude.ac/code/session_012arAwUuBQrpnMzJ36mEVBk
- tdd-gate: pass criterion is the task's TEST_FILES scope (matches
run-ci-tests.sh M19 P-14), not the full suite — contradicted
qwen-code-methodology rule 5 in the same system prompt
- tdd-protect: tests/ path described relative to project working dir
(AGENT_WORKING_DIR is /workspace/project in repo composites)
- qwen-code-methodology: task_complete tool DOES exist — rule 8 said
the opposite; pytest examples use -q
- airouter-code-tdd template: TT-39 runtime_overrides
(max_result_bytes 24KB, max_history_turns 16)
Claude-Session: https://claude.ai/code/session_012arAwUuBQrpnMzJ36mEVBk
All tasks dispatched to integration-direct/v1 were failing with
harness_missing_agent_repo_layer because dispatcher DSP-1 guard
requires agent-repo in the composite when AGENT_REPO_URL is set.
direct-push/v1 sets env vars (AGENT_SKIP_BRANCH_PUSH,
AGENT_EMPTY_DELIVERABLE_CHECK) that configure agent-repo's finalize.sh,
so agent-repo/v1 must precede it in the layer list.
MiniMax moved their Anthropic-compatible API from api.minimax.io to
api.minimaxi.chat. Update ANTHROPIC_BASE_URL in the minimax/v1 context
harness and corresponding comments in init.sh and the TODO note.
Model routing is unaffected: the API defaults unrecognised model names
(e.g. "minimax" from the registry slug) to MiniMax-M3 already.
Claude-Session: https://claude.ai/code/session_01XtWRazqRbMnMRa4iV3zbVx
New code-airouter-deepseekv4flash-tdd-repo/v1 composite (deepseek
methodology + the same tdd-protect/tdd-gate/python-dev/tdd-file-lock
layers as the Qwen TDD composite), plus a task-template that sets
reasoning_effort=max via the new extra_body passthrough (agent-runtimes
TT-38). Model/endpoint/model-registry entries already existed; only
the TDD-capable harness and the template were missing.
New context: hugo-content-workspace/v1
- init.sh: clones hugo-{customer}-content at HUGO_CONTENT_BRANCH and
hugo-{customer}-integration at main via gitea-ssh-accelerators
- finalize.sh: detects changes, commits content repo first then
integration repo; skips silently if no changes
- CLAUDE.md: instructs the agent to work in /workspace/content/
New composite: hugo-content-airouter/v1
- Combines airouter/v1 + gitea-ssh-accelerators/v1 + hugo-content-workspace/v1
- Label-gated to airouter dispatchers (ESO secret required)
Used by cms-proxy /ai/{customer}/draft endpoint.
Brings the framework CRS repo up to date with all content that was
living in agent-runtimes (local-dev fallback) but hadn't been promoted.
New composites: feature-delivery-loop, integration-direct, scaffolding-repo,
sonnet-impl-narrow, sonnet-manager, test-writing-repo
New contexts: integration/v1, scaffolding/v1, sonnet-manager/v1, z-ai/v1,
airouter/v1/bin (anthropic-compat-wrapper.sh), cp-harness/v1/init.sh
New task-templates: sonnet-integrator.yaml, workflow/* (17 typed workflow
task templates for the Epic 1 pipeline)
Updated: agent-repo/v1/finalize.sh — adds AR-38/F97 empty-deliverable audit
(SKIP_BRANCH_PUSH support, boilerplate-path filtering, ci_metadata.json flag)
Also adds MEMORY.md index and memory/ topic files for the framework repo.
Required by the Epic 1 trigger pipeline in agent-runtimes:
- spec-draft-opus@1 uses spec-writing-airouter-repo (concept→spec-draft workflow)
- review-spec-arch-opus@1 uses best-practices-airouter-repo (spec-draft→spec-reviewed-arch)
- scope-decompose-sonnet@1 uses spec-writing-airouter-repo (spec-reviewed-arch→scoped-tagged)
Both composites require the airouter label so only the airouter dispatcher claims them.
agent-repo/v1 init.sh creates /workspace/project/results/ and .gitignore,
making the directory non-empty before the integrator clone action runs.
git clone into a non-empty directory fails with 'already exists'.
Integrators don't need the agent-repo lifecycle — they clone origin/main
directly and push back. Remove agent-repo/v1 from the composite.
Copies code-sonnet-direct/v1 (composite) and direct-push/v1 (context)
from agent-runtimes local harnesses to the CRS-served framework repo.
These are required by the sonnet-integrator task template used in the
ralph coding loop. Without them, all integrator dispatches fail with
"harness not found in any concepts repo: code-sonnet-direct/v1".
Probes 7-9 (2026-05-08) all "succeeded" but with empty diffs because
AR-14a fell back to fork main: "WARNING: /workspace/reference/main/.git
not found; using fork main (may be stale)". The reference clone python
loop printed "Cloning git@..." but never "Cloned main successfully" —
no error visible either.
Most likely cause: full-history clone (post-3087ad7) is hitting some
silent failure mode (auth, fs perms, OOM, timeout). With capture_output=True
the git clone's own output was hidden.
Add explicit logging:
- python3 -u (unbuffered stdout)
- pre-loop "reference_branches loop: N entries"
- post-clone "clone returncode=N"
- always-print stdout (last 1KB) and stderr (last 1KB) from subprocess
- explicit chmod return code check
- post-clone .git existence assertion
Next probe will tell us EXACTLY what's failing.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Probe 7 (2026-05-08) finally surfaced the actual push error:
! [remote rejected] HEAD -> task-9dc266b5 (shallow update not allowed)
Cause: REFERENCE_BRANCHES handler clones with --depth 1, AR-14a then
fetches upstream-ref --depth 1. The agent's task branch is a single
commit on top of a single shallow commit — no ancestry visible. Gitea
rejects shallow pushes server-side.
Fix:
- Reference clone drops --depth 1 (full history)
- AR-14a fetch upstream-ref drops --depth 1 (full fetch from local-path)
- Cost: a few extra MB per task on tmpfs/PVC. Acceptable.
This unblocks AR-14a's upstream-seeding for the dogfood pipeline. The
agent's branch now has the full upstream history visible to gitea.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Probe 6 (2026-05-08) showed git push failing twice with exit 1 and
zero visible output — the previous form `if cmd 2>&1; then` redirected
git's stderr to stdout where the entrypoint's stderr-only log capture
missed it. CP-side log showed only the bash `set -x` trace, not the
actual git error message (e.g., "Permission denied (publickey)" or
"remote: pre-receive hook rejected").
Refactor the push retry loop:
- Capture output to PUSH_OUT via $() with `2>&1`
- Wrap in set +e/set -e to detect non-zero without aborting
- echo PUSH_OUT to stderr (where set -x trace also goes) so the
entrypoint's stderr capture sees it
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
