PR #592 — Finally, the CLAUDE.md Regeneration
The Broader Goal at the End of the Previous PR
PR #591 fixed the last known bug blocking synlynk sync --repair-sops --confirm from producing a correctly-formatted regeneration of the harness-managed instruction files. Nothing remained in the way.
Strategic Shifts in This PR
None — this is the capstone of a five-PR chain (#584, #588, #589, #591, #592) that started as a single "update CLAUDE.md" ask and kept growing one root-cause tooling bug at a time. No shortcuts were taken along the way: at no point was the harness-managed content hand-patched directly, even though that would have been faster at every single step.
What This PR Shipped
synlynk sync --repair-sops --confirm, run clean on a fresh branch off updated main:
- CLAUDE.md / GROK.md: refreshed the stale
## Capability-Based Task Allocationsection with the real per-repo routing table (generated from.synlynk/config.json, not synlynk's hardcoded default fleet assumptions) and the current GitHub-write-routing text — Grok-by-default, Agy's conditional headless capability (contingent on localsettings.jsonallow-rules, not reliably verifiable mid-task), Codex's sandboxed network block, and the #569 caveat that--requires-gh-write's token-stripping fallback isn't a hard identity guarantee. - GEMINI.md: had zero pre-existing SOP headers, so all six harness sections filled in fresh.
Formatting was manually verified section-by-section before committing — exactly one blank line at every boundary, headers rendering correctly — rather than trusting the tool blindly, given the newline bug found on the previous attempt (#591). This was a docs-only change, kept in its own PR per this project's own CLAUDE.md convention separating doc-only changes from code changes. Non-authoring review (Grok, via COMMENT fallback) confirmed the diff scope, table content against .synlynk/config.json, and formatting before merge.
Brainstorm Visuals Used
None.
What This Achieved on the Path to Autonomy
The instruction files every dispatched agent reads now accurately describe the fleet's real GitHub-write capability, instead of a stale "Grok only" rule that this same session's own evidence (Agy succeeding at PR #589's review/merge) had already contradicted. This closes a chain that started as a simple documentation update and, by insisting on root-causing rather than patching, surfaced and fixed four separate tooling bugs (#339's vacuous doctor checks, #339's own live-subprocess-in-hot-path side effect, #583's stale-detection gap, and #583's own newline-spacing regression) along the way.
Strategic Note: The Goal at the End of This PR
With doctor/probe trustworthy and the instruction files current, the fleet-parity audit (#332, #338, #340, #342, #347, #348, #419, #461) can resume on solid footing. Issue #573 (Agy can't call Stitch MCP tools headless) remains queued behind it.