← Back to Blog Oct 03, 2026 PR 1967

gh:#1924 PR5 — Claude, Local, and the End of the Strangler Pattern

The Broader Goal at the End of the Previous PR

PR4 (#1965) had just put Agy on the HarnessAdapter protocol, leaving one port outstanding. The goalpost coming into PR5 was explicit from #1965's own closing line: port Claude and local, then delete LegacyAdapter once nothing in the registry still depends on it — closing out gh:#1924 entirely.

Strategic Shifts in This PR

The plan said "delete LegacyAdapter." The PR that shipped does not. That's not scope creep — it's a correction the first version of this PR got wrong and CI caught immediately.

dispatch_agent()'s original fallback was:

try:
    adapter = get_adapter(request.agent)
except KeyError:
    from synlynk.harness_adapters.legacy import LegacyAdapter
    adapter = LegacyAdapter(agent=request.agent)

PR5's first commit replaced this with a bare get_adapter(request.agent) call, on the assumption that once all five core harnesses had adapters, nothing could still hit the KeyError branch. That assumption was wrong: muse (Meta Muse, Milestone v0.20.0 Cluster D, #1508) is a sixth harness that was never in scope for this five-harness migration, and it was never going to get its own adapter here. All four CI matrix jobs failed identically — tests/test_muse_harness.py::test_muse_dispatch_flags_and_command_construction raising KeyError: 'muse' straight out of registry.py.

The fix restored the fallback exactly as it was, and un-deleted legacy.py and test_legacy_adapter.py, which the same first commit had also removed. LegacyAdapter is now explicitly permanent infrastructure — the generic landing spot for any harness name outside this migration's five-harness scope, not a placeholder scheduled for deletion once the migration finished. The strangler pattern strangled the conditionals, not the concept of a fallback.

What This PR Shipped

  • ClaudeAdapter and LocalAdapter implementing the HarnessAdapter protocol, matching the translate_permissions() / build_cmd() / parse_output() / classify_failure() / resolve_model() shape already established by CodexAdapter, GrokAdapter, and AgyAdapter.
  • LocalAdapter fails closed. aider (the local/oMLX harness) has no mechanism to enforce a read-only or file-scope restriction, so translate_permissions() raises PermissionEnforcementError the moment any permission is requested, rather than silently granting full read/write access it can't actually bound. Architect review specifically verified this is enforced, not just asserted in a test.
  • registry.py now registers exactly five harnesses — codex, grok, agy, claude, local — and nothing else. muse (and any future sixth harness) stays on the LegacyAdapter fallback by design.

Caught in Review

Two things surfaced after the muse regression was fixed, both before merge:

  1. A real merge conflict, not a stale GitHub cache artifact. PR #1962 landed on main mid-flight, also touching README's test-count badge/hero line — the same CI gate that had already bitten PR4 and PR4's predecessor. gh pr update-branch correctly refused (it only handles non-conflicting fast-forwards), so the fix was a manual merge, re-running pytest --collect-only -q on the merged branch to get the true count (3764) rather than taking either side's stale number, and patching both the badge and the hero line to match.
  2. Branch protection caught a second gap after merge, on the next PR in the queue. #1966 (this same session's blog-batch PR) merged to main just ahead of #1967, which put #1967's branch BEHIND again. One gh pr update-branch cycle, a CI re-run (all six checks green), and the merge went through clean — exactly the "at most 2 update-branch → CI-wait cycles" discipline this repo's PR Review Discipline calls for.

Architect review (non-author, routed here rather than qa because this PR is architecture-impacting — it closes a cross-cutting migration) verified the restored fallback, confirmed legacy.py/test_legacy_adapter.py exist, confirmed the registry's exact five-harness membership, and ran the full suite: 3,759 passed, 4 skipped, 1 sandbox-only logger failure that passes with an isolated writable HOME. Approved.

What Was Achieved Toward the Long-Arc Goal

All five harnesses — Codex, Grok, Agy, Claude, local — now route through dedicated HarnessAdapter implementations instead of the scattered if/elif branching that used to live directly in dispatch_agent(). Each of the four ports in this series caught something real before it shipped: Grok's port surfaced a stale capability-routing bug; Agy's port caught a missing lifecycle-telemetry population; Claude/local's port caught its own regression against a harness explicitly outside the migration's scope. The pattern did exactly what it was brought in to do — turn harness-specific quirks into isolated, testable, single-adapter concerns instead of silent gaps in a shared conditional.

gh:#1924 is closed.

The New Goalpost

The adapter migration itself is done; what it leaves behind is a cost-anomaly thread worth watching. Three of the four dispatch jobs that reviewed and fixed this PR cluster (#1966 and #1967 together) ran 6x–130x over their own preflight cost estimates, two of them tripping critical cost-inflation sentinels ($14.69 and $21.34 on jobs whose estimates were under a dollar). Filed as #1969 rather than folded into the five prior cost-inflation issues that closed without a hard cap — this one has harder evidence (same session, same PR cluster, consecutive jobs) and a specific suspected mechanism (conflict-resolution retries inside a dispatch possibly not respecting --context-mode's intent to minimize injected context). That's the next thread to pull, not another adapter port.

Refs #1924, #1969.

← Back to all posts