URLs, mount paths, and LAN host bindings parameterized via env or relative paths
so the repo stands up from a clean clone anywhere. Drop cross-codebase refs
("mirrors clawdforge's pattern"), Sulkta-Coop client/merchant test fixtures,
and audit-changelog scaffolding from comments. README terser, technical content
preserved.
The 4 patcher-fired-but-malformed_response failures showed extract_diff_json
was too strict: it required {"diff": "..."} as the top-level JSON shape
with at most 1 brace nesting depth (regex-based). Real model output
varies more.
Now handles:
1. Bare JSON {"diff", "explanation", "confidence"}
2. Fenced JSON: ```json {…} ```
3. Fenced diff + prose: ```diff …unified diff… ``` + loose explanation
4. Bare unified diff (no JSON wrapper, no fence)
5. JSON with deeply-nested {} inside the diff string (struct literals,
function bodies)
Fixes:
- Replaced regex-based balanced-{} matcher (capped at depth 1) with a
string-aware depth-tracking generator that handles arbitrary nesting
+ skips brace chars inside JSON string literals
- Walk all fenced blocks not just the first; recognize ```diff and
```patch language tags
- Fall back to fenced-diff-with-prose construction when no JSON form
matches — synthetic payload with surrounding text as explanation
- Final fallback for bare unified diffs (no fence, no wrapper) using a
simple line-prefix detector
- Normalize alternate keys (patch, content, diff_text → diff)
- Always set confidence (defaults to medium when absent, low for bare
diffs that have no model commentary)
Tests: 16 → 20 (5 new shape coverage tests). All green.
Two bugs caught by the first real concurrent dogfood (15 SDK build jobs
queued at once for clawdforge):
1. Concurrent fetch race — multiple jobs from same project all called
`git fetch +refs/heads/*:refs/heads/*` simultaneously. Git refuses
to fetch into a local branch ref that another worktree has checked
out, so jobs after the first failed fast with:
fatal: refusing to fetch into branch 'refs/heads/main' checked
out at '/workspace/clawdforge/<other-job-id>'
2. Worktree-from-local-branch — `worktree add ... main` reserved the
local branch ref to the worktree, blocking subsequent fetches even
when the lock above wasn't held.
Fix:
- Per-project asyncio.Lock around the materialize() body (clone +
fetch + worktree-add). Different projects still parallelize; same
project serializes through the cache dir.
- Fetch into remote-tracking refs only:
+refs/heads/*:refs/remotes/origin/*
Local refs/heads/* are never written, so no worktree can hold them.
- worktree add uses `--detach origin/<branch>` so each worktree is at
the remote-tracking ref. Multiple detached worktrees share the same
remote ref without conflict.
The recipe itself runs outside the lock, so concurrency=4 still
parallelizes the actual build/test work — only the brief
materialize step (clone or fetch + worktree create) serializes.
Tests still 6/6 green on test_runner.py (uses _StubWorkspace, not the
real WorkspaceManager, so it's unaffected — the real exercise is the
live re-queue against fixed code).
Code-work prompts (read CVE/lint context, draft a unified diff that
verifies cleanly against the failing recipe) reward Opus's longer
context + careful reasoning. Cauldron-style high-frequency Sonnet calls
are unaffected — this only changes what crafting-table's patcher asks
clawdforge to use for its drafted-patch sessions.
Pairs with clawdforge dbbead2 which adds the optional `model` field on
POST /sessions and propagates ANTHROPIC_MODEL into the acpx subprocess
env on both create + turn.
Override knob: CRAFTING_PATCHER_MODEL env (e.g. "sonnet" if cost > quality).
Tests: 16/16 patcher tests still green.
