idkmesh

Pull Request Triage — 2026-08-28

Repository: MSKazemi/idkmesh
Initial open PRs: 26
Final open PRs: 1
Final open PR: #91, intentionally draft for separate human review.

Purpose

Reduce integration debt without optimizing for raw merge/closure count. Preserve evidence, negative results, project memory, safety boundaries, and one canonical implementation per responsibility.

Maintainer rules used

  1. Review the exact diff and exact-head CI, not only the PR description.
  2. Integrate in dependency order.
  3. Prefer clean current-main replacement PRs over stale ancestry/rebase gymnastics.
  4. Keep evidence/runtime gates binding even when unit tests are green.
  5. Preserve failed experiments and runtime failures as evidence.
  6. Close duplicate controllers and duplicate acceptance/evidence paths once one canonical path exists.
  7. Preserve conversation records before closing stale archive-only PRs.
  8. Split unrelated research/product/memory concerns before review-ready status.
  9. Do not let automation infer approval or merge authority from its own success.
  10. Keep only genuinely blocked work in draft.
  11. Treat live repository state as concurrent: re-check before every irreversible merge/closure and reconcile races publicly.

Direct integrations from this triage

#103 — verifier output authority

Merged after exact two-file review and green Phase 0, Evaluator Plan Binding, and Evolution checks.

Result: verifier/evaluator-generated files are confined to results/; canonical repository paths such as README.md cannot be targeted as generated evidence.

#100 — R4 verified stigmergic routing harness

Merged as a small green research leaf. Unverified activity cannot positively reinforce the routing signal and the research harness has no acceptance/governance authority.

#111 — patch evaluator evidence completeness

Merged on exact head c01f768f5e6efaaac9688a8c2243ce27e1b20d4d after its dedicated Evaluator Plan Binding check, Phase 0, and broader suite passed.

The follow-up closes three concrete evidence gaps:

Unsupported/ambiguous patch shapes fail closed. No patch application, candidate execution, secrets/network access, or merge authority was added.


Important concurrent integrations reconciled during triage

The repository evolved rapidly while this queue was being reviewed. Useful work was not duplicated when another clean convergence PR landed first.

Canonical integrations observed during the window include:

These integrations changed the value and dependency status of many older PRs, so stale parallel branches were not preserved merely because they existed first.


Closed / superseded / deliberately unmerged work

#101 — one-shot R4 result-generation workflow

Closed because the PR description claimed frozen result artifacts while the actual diff contained only a write-capable self-pushing workflow. Evidence should itself be reviewable; a one-shot workflow need not land on main merely to manufacture it.

#64 — parallel ONE continuous controller

Closed as a second write-capable controller/state/label/actuator system. Useful ideas such as one bounded next action and carrying-capacity pressure belong in the canonical ACE stack instead of a parallel autonomous controller.

#50 / #94 — stale archive-only PRs

Conversation records were copied to current main before closure, removing merge debt without deleting project memory.

#44 — obsolete cumulative-load population experiment

Closed as historical research evidence after the recoverable live-open-work capacity model replaced the cumulative historical-load assumption.

#43 — GitHub Reflex Observatory draft

Closed unmerged after review identified a verification-process gap: a newly introduced pull_request_target workflow was not evidence that PR-head collector/tests had actually executed. The read-only architecture remains useful, but a future observatory should provide a trusted PR-head test surface before integration.

#105 — stale evaluator stack

Closed as superseded by clean current-main replacement #107, which merged.

#108 — exact-head controlled-runtime acceptance harness

Closed unmerged after completing its evidence purpose. It was intentionally candidate-bound instrumentation, so landing it on main would create stale one-shot maintenance surface.

Its failed and passing runs remain public evidence.

#114 — duplicate/obsolete negative runtime matrix

Closed because it froze rejected old worker SHA d638a2f... and reimplemented tests already completed successfully by #108/#37 against the current worker SHA. A second frozen-candidate acceptance implementation would fragment the evidence path.

Other stale/parallel lanes

Older PRs including #36, #40, #55, #56, #59, #63, and #89 were closed or superseded as the canonical implementation/evidence stack converged. Their public discussions remain provenance; unique ideas should return only as bounded deltas on accepted surfaces.


Real-worker runtime evidence: negative result -> correction -> PASS

The most important triage outcome was not a merge. It was a real falsification.

Failed frozen candidate

Old PR #91 candidate:

d638a2f78e4a89353b98e91052233e365f56f90a

#108 run 33185574177 reached the real Docker runtime and correctly resolved immutable image evidence, but the positive task failed with a Python SyntaxError.

Root cause: JSON decoding turned newline escapes into literal newlines inside a single-quoted python -c string.

The repository did not weaken the test or call this success.

Correction

The canonical fixture switched to an escape-free chr(10) construction and Node CI gained a regression test that parses and compiles the decoded task command.

Final exact worker candidate:

520ad2c9aa5825476de4957da4702d6823f4edb3

Prerequisite exact-head CI:

Controlled runtime PASS

#108 run 33186111350 passed the full positive + A–E2 matrix for exact head 520ad2c9....

Positive evidence included:

Negative cases all failed closed:

Issue #37 is completed for exact head 520ad2c9....

The evidence explicitly records worker_acceptance_authority: false.


Real node -> independent evaluator proof

PR #113 subsequently merged the first direct proof that the real accepted node bundle can be passed through the canonical hardened EvaluatorPlan/unified-diff verifier to a canonical VerificationResult.

The worker checkout and evaluator checkout remain separate, the worker produces its bundle before the plan is constructed, and the metadata-only verifier executes no candidate code.

A verifier recommendation remains decision support, not merge authority.

This materially reduces the product critical path: the missing problem is no longer “can a real node produce inspectable evidence and can the independent verifier consume it?” Both now have direct evidence.


ACE convergence state

The canonical ACE Phase-A evidence/control stack now includes:

#106 observer/exposure inventory
 + #48 causal lineage
 + #104 live-open-work capacity
 + #68 shadow strategy controller
 + #98 privileged-workflow safety boundary
 + #112 external Phase-B activation gate

The Phase-B gate remains fail-closed even when review capacity is healthy. Current activation is blocked by real missing authority/evidence, notably protected integration and independently verified descendant evidence.

Healthy capacity is explicitly not permission.


Final PR queue

#91 — canonical local real worker

Only open PR at the final triage snapshot.

State: draft, mergeable.

Exact head:

520ad2c9aa5825476de4957da4702d6823f4edb3

Completed evidence:

Remaining blocker:

A separate human/reviewer must inspect the exact-head CI/runtime/evaluator evidence before #91 leaves draft.

The same repository owner, worker, CI harness, or automation must not manufacture its own independent approval.

PR #91 is labeled help wanted for that remaining review.


Canonical next product path

#91 independent human evidence review
 -> integrate canonical local worker
 -> real worker adapter behind merged two-attempt orchestration (#78)
 -> merged non-selecting Evidence Report/replay (#88)
 -> first complete real replayable Verified Swarm Runner run (#16)
 -> bounded real-task experiment (#2/#30)

Do not open another canonical worker/verifier/orchestrator/report implementation while this path exists.


Definition of a healthy merge queue

Prefer:

A useful direction for integration health remains:

IntegrationYield =
  verified durable changes merged
  ---------------------------------------------
  reviewer attention + conflict debt + rework

with penalties for stale review-ready PRs, duplicated responsibility, description/diff mismatch, and unverified privileged workflows.

Immediate maintainer rule

With the queue reduced from 26 open PRs to 1, stop optimizing for more closures/merges. The next high-value action is the independent review of #91, not another architectural/controller PR.