PRKrew, building PRKrew.

Issue 326, filed by PRKrew's own review round on the budget gate work. Ranked low and picked up as cleanup of the grant path.

Issue 305, filed by PRKrew's review round against the repo's own maintainability guideline on unit complexity.

Issue 340, another finding from PRKrew's review round on the launcher work, the third refinement of the same promise.

Issue 341, filed by PRKrew's review round. No single source fix exists across Dart and XML, so the answer is a guard that fails on drift.

Issue 306, filed by PRKrew's review round on the iterate work loss fix, aimed at the same adoption logic as the helper extraction.

Issue 342, filed by PRKrew's own review round while checking an earlier mobile fix. Not reachable on the shipped Android build, so it was ranked low and picked up as cleanup.

Issue 343, filed by PRKrew's review round on the earlier Open Tailscale fix. Same class of dead tap as that fix, just in a rarer device state.

Issue 328, filed by PRKrew's review round after reading a comment that promised more than the code delivered. On local SQLite the gap was milliseconds, but the contract was wrong.

Issue 344, filed by PRKrew’s own review round on a mobile pull request. Honest detail: the review could not compile the Kotlin or read the pinned plugin version, so it filed the finding as a plausible toolchain risk rather than a proven break.

Issue 332, filed by PRKrew’s own review round as the minor finding on the OTEL tracing pull request. It named the exact line, the corrupted URL it produces and the invariant the fix should restore.

Issue 315, filed by PRKrew's own review round on an earlier pull request as a minor finding. It named the exact line and the exact shape of the correction, so the pipeline had a very grounded starting point.

Issue 307, filed by PRKrew itself. It recognised the diverged case as the third member of a lost work family the codebase already handles, next to empty steps failing outright and the lost work tripwire, and asked for the same treatment.

Issue 333, filed by PRKrew itself. Honest detail: no trigger exists today, every send is already wrapped. The fix makes the exporter's never fails a run contract structural instead of aspirational.

Issue 316, filed by PRKrew's own review round. An off by one in prose rather than in code, and the existing test was guarding the bug instead of catching it.

Issue 317, filed by PRKrew's own review round, noticed the event had a feature prefix in its name but nothing on it said which feature was oversized. The identifiers were already in scope at the emit site, just never passed through.

Issue 334, filed by PRKrew's review round of the tracing work, pointed out that the promise of standard variable names being picked up for free breaks the moment the most common paste source uses percent-encoding.

Issue 318, filed by PRKrew's review round, spotted that the admission estimate was re-priced through calibration while the overhead note used raw means, so the human deciding at the gate compared numbers on different scales.

Issue 335, filed by PRKrew itself during a review of the tracing work, warned that APM panels faceting on the well known operation names would show the root of every trace as an unclassified operation.

The intake agent picked up the second half of Kevin's sandbox issue after pull request 281 shipped the policy model, the settings card and the run detail surface.

The Director picked this on its evening tick after the numbers showed 53 of 79 recent features stalled, making dropped review triggers the single largest drain on autonomy.

The Director picked this on its evening tick as a second, deliberately smaller attempt. The first try reworked the budget engine, spent 47 dollars and produced no pull request.

The Director picked this on its evening tick: a red app suite means every merge passes a quality gate nobody can trust, which quietly undermines everything else.

Kevin filed issue 146 after the ecosystem watch flagged sandbox hardening as a widening gap: langflow shipped a microVM execution backend and dify fronts its runs with an egress proxy, while our container still ran without boundary flags.

End-to-end run 15 drove a real issue through the local lane on a fresh install. Every step passed, yet the feature ended as passed with no pull request and no trace of the skip. It was the third reproduction of this silent-skip class.

On the machine PRKrew runs on, CI-failed banners fired on repos that have no CI at all and were already being ignored. An alert everyone ignores is worse than none, because the one real failure would slip through with it.

End-to-end run 14 on a fresh database exposed the silent fallback: a first-time user got a bad plan with no explanation. PRKrew filed issue 309 as a concrete first slice, replacing a broader open issue.

The ecosystem watch flagged standard GenAI tracing as the fastest rising gap it tracks: five surveyed repos added it within one month, n8n and langflow among them. That signal became issue 286.

Direct work from a hands on session, found while testing the first phone build from the new Play pipeline. Not a pipeline run, a human noticed the dead button.

PRKrew filed issue 308 about itself, straight from the findings of end to end run 14 on a fresh install: the budget cap was crossed blind at 102 percent, and the first signal the user got was the stop itself.

PRKrew filed issue 310 about itself from the run 14 cost statistics: over decomposition had become the dominant cost driver on small issues.

Direct work from a hands on session after Kevin asked for it plainly: setting up phone access should happen from PRKrew, not with separate commands in a terminal.

Direct work from a hands on session, not a pipeline run. The mobile pairing app needed a way to reach phones beyond the dev PC, and a Play release channel is that road.

PRKrew filed issue 279 about itself after end to end run 12 left feature 42 wedged in exactly this loop, retrying a step its own duplicate gate kept refusing.

Issue 148 came from the ecosystem overview, the routine that watches what comparable tools ship. Installable, permission scoped agent catalogs are becoming a standard capability, and PRKrew wanted its own honest version.

Kevin found it on the director PC: two review fix runs reported nothing pushed, while their finished commits sat as dangling objects in the clone. The work was rescued by hand, then the root cause was traced in the audit trail.

PRKrew filed issue 278 about itself after end to end run 12 failed. The root cause was a skill that existed in the agent’s configuration but never made it into the run’s worktree.

PRKrew filed issue 289 itself as a re scoped second attempt. The first try was cut too wide and failed, so this slice was confined to the single decision point, mirroring the pattern that made budget auto resume succeed.

This came from a direct question in a desktop session. Kevin asked whether the cron agents carry the ISO 25010 maintainability skill, and the honest answer was that none of them did.

PRKrew picked this itself. The Director filed issue 257 as a narrower first slice after an earlier attempt at auto resume failed, budget parks caused 14 stalls in the two weeks before.

From the ecosystem overview scan. Six comparable projects ship run level scoring, and this gap sat at the top of the ranking for eleven consecutive scans before it was filed as issue 145.

From the ecosystem overview scan. Comparable tools just promoted execution hooks from a callback into a stable contract with failure semantics, and issue 149 asked PRKrew to offer the same seam.

No issue behind this one. The dropdown got in the way during daily use of the board, so it was rebuilt directly in a desktop working session rather than through the pipeline.

The first of two guards Kevin asked for directly during the 2026-09-06 stability audit, right after the audit train itself had merged.

The second guard deferred from the 2026-09-06 stability audit, requested by Kevin together with the wedge guard.

The pipeline's own test run found it. A rogue agent PR made the step think nothing happened, and the checkpoint system turned one failure into a loop the halt message could not explain.

A review round on an earlier PR spotted the collapse in the state fusion, and Kevin filed it as issue 260. The live proof was already there: the open Director PRs had not moved since September 3.

Part of the September 6 stability audit, done before more routines move inside PRKrew. The eight legacy PRs alone would have cost 24 model calls on the first tick for reviews that already existed.

All three defects were observed live on the machine that runs the Director: 1808 skipped runs in one day from the typo, duplicate reviews from overlapping ticks, and a review agent that showed zero runs because the build was older than its handler.

The Director PC was running a September 2 build while master already carried the new PR review handler. Every hourly session pulled the repo first, which silently defused the rebuild check in the update script.

Direct follow-up to the review round that moved inside PRKrew earlier the same morning. Shipping the handler was not enough, every deployment also needs the two agents that trigger it.

Kevin watched the external review routine and the fix loop chase each other on one PR for more than a day, then filed four issues describing the guardrails the routine was missing.

This grew out of the hourly review routine set up on September 2. The routine could leave a verdict on a PR, but nothing in the pipeline read it, so every fix still needed a human to pass the findings along.

The Director picked this itself on day 8. Its two bigger picks were parked waiting for approval, so with the budget point it had left it chose a first run papercut on the install to first PR path.

The header checked for a status called done, but the schema only knows pending, running, passed, failed, blocked and planned. The check could never be true, so the elapsed label never switched.

Kevin set the target in a desktop working session on September 1 and it was written down the same day. Not a pipeline run, this is the map the pipeline follows.

A desktop working session, not the pipeline: the Claude 5 family had shipped and PRKrew's pickers still stopped at Opus 4.8.

A live-test finding: the August 31 end-to-end run watched the planner notice a contradiction in its own plan, implement the stale text anyway, and re-plan work already sitting in four open pull requests.

Found during an autonomy review: the events and even the database columns for this loop had been reserved for months, and a red PR just sat there. No re-run, no inbox item.

The failure taxonomy said everything was broken while the guard was doing its job. The same autonomy review that produced the other gap issues flagged this one.

Three unrelated-looking test failures on the development PC all traced to twin processes sharing one worktree directory: a commit that had already been made, a delete blocked by another process, a worktree registered to someone else's clone.

Confirmed twice in the end-to-end check series (August 26 and August 30): the poll correctly selected only one project's repos, then handed the feature to the repo's home project instead.

Measured in the end-to-end series: fourteen of fourteen recent failures were budget halts, and the admission estimate ran roughly 28% low. An estimate that is systematically low is worse than none.

On August 30 and 31 the Director skipped over its monthly cap and produced nothing. Correct guard behaviour, but the only trace was a line in its decision history.

The publish race had failed a genuinely delivered run twice in the end-to-end series before it became an issue, the second time on the run with the best deliverable yet.

Spotted while adding Fable 5 to the catalog and verified against the provider's current price table: every current Opus release bills a third of what the table claimed.

A desktop working session: the CLI could already run Fable, the product could not offer it.

The step-level semantics were rescued from a pull request the pipeline-integrity train had closed as superseded, and rebuilt on the current commit design.

After nine straight runs died at “budget limit reached”, the Director filed the enforcement half of its own budget problem: runs were being killed over money nobody had actually spent.

Two days after a schema migration bricked every installed database, the incident became a guard: an upgrade has to be provably survivable before it ships.

PRKrew promises honesty about cost, but a new user picking a budget had nothing to go on: the data sat in every run record, unpublished.

The Director’s multi-pick worked the migration-guard issue in two parallel lanes, and both produced a complete implementation.

The second parallel lane on the cost report: same issue, independent branch, its own generator.

The enforcement fix itself had already merged in the repair train. What remained of the issue was visibility: a budget halt has to appear where a human actually looks.

A repeating failure on the director PC: a service killed mid-planning left features showing a busy placeholder graph forever, and only a human noticing by eye ever rescued them.

A forced re-run once crashed the service and left no logs at all, no stack, no last action, and the root cause was never found. That diagnosis-by-guesswork is what this removes.

The board offered Run and Accept on plans that did not exist yet, and Start planning on a project with no repos returned a raw 422: buttons lying about state the user could not act on.

The admission gate shipped with 44 tests, but two of the five ways a run can start had no test at all: exactly the entry points a route-level check would have missed.

“Add budget and continue” existed, but two paths could still consume the inbox item while the run stayed halted, leaving the human at exactly the dead end the issue was meant to remove.

Every recent mid-flight halt was predictable for free before the first token was spent, so the Director filed the admission side of its budget problem.

Run failures over two weeks were dominated by a single reason: budget limit reached. The Director filed the recovery half: a stop should be a pause, not a death.

Reviewing a similar open source tool showed what the mobile idea is really for: saying yes or no to a plan from your phone. That need does not require a whole app.

Found live, re-verifying a rescued feature: the shallow workspace clone had drifted onto a disjoint history island the moment master advanced, so the chain publish silently skipped, caught again by the empty-run backstop.

In the live rerun two features delivered real PRs, but a chain’s final step changed no files of its own, so the publisher saw “no changes” and skipped publishing the whole chain’s work.

Last car of the four-PR repair train: enforcement, durability and delivery were fixed. Admission was still missing, and every mispriced run kept burning real money to discover its own halt.

Three features had reached “passed” with zero pull requests. QA’s own not-ship-ready verdict was recorded as evidence and then ignored. A run that delivers nothing must say so.

The root of the empty-PR mystery: step work lived uncommitted in a shared worktree, and one mid-run park force-deleted finished implement work. After that, QA happily passed on a pristine tree.

Nine straight self-dev runs died on “budget limit reached” at impossible numbers: $15.14 of $15.00 on 41k tokens is $367 per million. Something was counting money nobody spent.

The daily ecosystem overview filed its top feature gaps as issues nobody handled, while the Director took exactly one item a day. Reliability always outranked them. The human call: keep the Director as decision-maker, let it take several items by complexity.

The cron-agent dialog’s core flow, picking the project and repositories for a GitHub-intake cron, was unreliable: unscoped organization walks, silently swallowed failures, stale ids re-posted on save.

A self-chosen onboarding gap: a fresh machine showed an empty Runtimes page and expected you to already know which CLIs to install, and how.

The Director’s daily issue lived only on the local board. Filing it on GitHub makes the self-dev loop publicly inspectable and lets the normal intake path claim it.

The Director’s persona was a hard-coded constant. Making it a real agent row means the decision-maker can be edited like any other agent.

The device-pairing API existed, but pairing a phone still needed a curl command, flagged as the open gap when the mobile companion shipped.

On its first supervised tick the Director evaluated five candidate features and scored this one highest-leverage: the last hand-run command between a fresh clone and a setup that survives reboots. Highest onboarding win, lowest risk: size M, risk low, by its own assessment.