The Claude Part_2 domain workers (Stage_B_B1..B5) had use_tools: [] and
relied entirely on preflight injecting stage1_tmp/task_c_bo/domain_slices/
B{n}.json. In the failing run the executed A0 wrote full-name slices
(B1_Money_Successor.json ...) while preflight expected short names
(B1.json), so nothing was injected and the tool-less workers emitted
tool-call syntax as plain text (<mcp_tool_call>/<function=read_file>) and
produced FAILED/empty output.
The current v.7 A0 already writes short names matching preflight, so the
naming mismatch itself is resolved in this file. This change adds
use_tools: ['localdocs'] (already declared on the stage) so the workers can
read their slice directly if preflight ever misses again — matching the
proven Codex worker design and removing the single point of failure.
Note: domain slices are 255-300KB (~65-90K tokens); llm_bridge caps tool/
injected content at 50K tokens, so slice compaction in A0 is still needed
for full-fidelity output quality.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Task_C_BO_R0_seed_reducer_exception_planner raised RuntimeError (exit 1)
whenever status became BLOCKED, and BLOCKED was triggered by validation
findings that are inherently fragile because they require an LLM (Stage B
workers) to echo values verbatim or stay perfectly in scope:
- slice_digest_sha256 echo mismatch (all 5 domains)
- run_fingerprint echo mismatch
- source meeting-clause refs outside slice/global (B5, 42 findings)
A decrypted postb_seed_ledger.json from the 2026-07-20 run confirmed all
47 blocking failures came from exactly these three check classes (no BLOCK
severity reviews / no budget overflow). Downgrade them from fatal
`failures` to non-fatal `reviews` (candidates preserved, routed to human
review). Genuine contract violations (schema/domain mismatch, worker
FAILED, candidate_ref sequence, forbidden keys, invalid enum) and the
deterministic A0-artifact integrity raises are kept fatal.
Gate simulation on the confirmed inputs: BLOCKED/exit-1 -> READY_WITH_REVIEW,
fan-out restored so R1 exception adjudication can run.
Note: this unblocks the pipeline and flags the issues; the upstream root
cause (Stage B tool-content truncated 66-78K -> 50K tokens) still needs a
slice-compaction fix for output quality.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
user_id is now a required workflow_dispatch input (no default) and the
runner validates it is non-empty before doing anything. Removed the
'jsahn' default from the script arg and replaced jsahn with <user_id>
placeholders throughout TEST_WORKFLOW.md so the workflow is not pinned
to one account.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Created new files for Stage 1 evaluation and revision strategy based on GPT and Claude recommendations.
- Added detailed revision plans addressing runtime issues in LES2 and optimizing workflows for Stage 2.
- Updated existing Stage 1 to Stage 2 transition prompts with additional search results and completion indicators.
- Enhanced clarity and structure in the documentation for better usability and understanding.
- Documented the workflow for running Agent YAML using Gitea Actions.
- Included setup instructions for .env file with Gitea token.
- Provided three methods for triggering the workflow: via Gitea API, Gitea web UI, and local script execution.
- Added workflow input reference and prerequisites for server infrastructure.
- Included troubleshooting section for common issues encountered during execution.
Add workspace_name input; the runner resolves it to a workspace_id through
GET {api_base}/workspaces/lookup?user_id=... before uploading the agent.
workspace_id remains as an explicit override. Name matching is NFC-normalized
(handles Korean NFD/NFC) with a case-insensitive fallback, and a no-match
error lists the available workspace names. Requires one of name/id.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Gitea Actions only supports the v3 artifact protocol; upload-artifact@v4
uses the @actions/artifact v2 API which Gitea (identifying as GHES) rejects
with GHESNotSupportedError. The agent run step itself succeeds; only the
artifact upload was failing.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
localhost:8800 fails because act_runner job containers run in their own
network namespace. Point api_base at the backend container on the shared
docker network (http://agent-backend:8000). Requires the runner's
config.yaml container.network to be set to the backend's docker network.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The public URL (legalpoc.eroomai.com) sits behind a Google OAuth proxy,
so CI requests get 302-redirected to the login page. Point api_base at the
internal backend (http://localhost:8800, served at root without /api prefix)
which the act_runner reaches on the same host. Documents host.docker.internal
and Tailscale fallbacks for containerized runners.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- .gitignore: add .serena/ and .DS_Store (both match at any depth)
- untrack 4 .serena files and 155 .DS_Store files (kept on disk)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
jhogyu
jsahnandClaude Opus 4.8