Skip to content

20 indexed bugs & failure modes

Agentic coding bug index

Historical defects and preventive patterns recovered from reviewed work, plus current failures caught by acceptance checks. Each record qualifies its observation and repair evidence; production health requires separate verification.

20 of 20 records

Bug and evidence index JSON

BUG-0001 · Post-Once · Concurrency

Stale parent and orphan children consume concurrency

Historical repair reported

Evidence scope: 9 August reported recovery and natural-cycle checks; not a fresh host observation.

Inspect mechanism, repair and regression check
Trigger
A parent attempt dies while durable children retain running state.
Symptom
New work stays capacity-blocked after the real worker is gone.
Root cause / qualification
Occupancy follows raw workflow status rather than canonical live attempts and leases.
Consequence
The scheduler can stop admitting useful work indefinitely.
Repair pattern
Reconcile attempts, leases and effects; recover only proven dead, effect-free targets.
Regression check
Recover orphan children; reject live or effect-ambiguous recovery; verify zero unrelated changes.
Next acceptance
Read current occupancy and effect receipts before recovery.

BEV-0001 · Dated acceptance report · indexed 2026-10-09

Inspect supporting sourceRelated skill: Stale Lease Recovery and Single Flight

BUG-0002 · Post-Once · Concurrency

Time-bucket idempotency permits overlapping monitors

Historical repair reported

Evidence scope: Recovered historical single-flight procedure; no current scheduler mutation executed.

Inspect mechanism, repair and regression check
Trigger
Equivalent periodic work enters a later time bucket while a prior attempt is live.
Symptom
Two equivalent monitors occupy the same lane.
Root cause / qualification
Time-bucket deduplication lacks a separate live-equivalence admission gate.
Consequence
Duplicate work competes for capacity and may repeat consequences.
Repair pattern
Add single flight keyed by workflow, lane, version and canonical live state.
Regression check
Equivalent concurrent ingress admits once; distinct lane/version admits; later expiry permits admission.
Next acceptance
Run bounded concurrent ingress acceptance on the deployed admission path.

BEV-0002 · Dated acceptance report · indexed 2026-10-09

Inspect supporting sourceRelated skill: Stale Lease Recovery and Single Flight

BUG-0003 · Post-Once · Measurement

Publication evidence becomes a measured baseline

Historical repair reported

Evidence scope: 13–14 September observations, not current campaign performance.

Inspect mechanism, repair and regression check
Trigger
A predecessor has verified publication but no comparable performance snapshot.
Symptom
A challenger is admitted before a measurable baseline exists.
Root cause / qualification
Effect equivalence and measurement maturity are collapsed into one gate.
Consequence
Learning and experiment claims become unsupported.
Repair pattern
Require publication equivalence, comparable measurement and challenger spacing independently.
Regression check
Published but unmeasured predecessor defers; mature comparable evidence permits admission.
Next acceptance
Read fresh metrics and test missing versus mature snapshots.

BEV-0003 · Dated acceptance report · indexed 2026-10-09

Inspect supporting sourceRelated skill: Measurement-Gated Learning Admission

BUG-0004 · Cross-project · Measurement

Unavailable metrics become zero performance

Preventive failure mode

Evidence scope: Grounded in recovered empty/unavailable evidence; not a newly reproduced defect.

Inspect mechanism, repair and regression check
Trigger
Consent, provider availability or low exposure prevents measurement.
Symptom
An agent interprets empty evidence as zero activity.
Root cause / qualification
Missing, blocked, immature and measured-zero states are collapsed.
Consequence
False business conclusions drive unjustified optimisation.
Repair pattern
Preserve explicit unavailable, immature and measured-zero states.
Regression check
Empty connector proves connection only; zero requires an observed measured interval.
Next acceptance
Check current unavailable-state rendering and reporting.

BEV-0004 · Reviewed preventive pattern · indexed 2026-10-09

Inspect supporting sourceRelated skill: Measurement-Gated Learning Admission

BUG-0005 · Post-Once console · UI and contracts

Guide interaction leaves a partial no-snapshot dashboard

Historical repair reported

Evidence scope: Recovered October dashboard repair; not independently repeated live here.

Inspect mechanism, repair and regression check
Trigger
A streamed interaction changes dashboard state during snapshot updates.
Symptom
Attention information disappears and sections partly render no-snapshot.
Root cause / qualification
The recovered repair concerns inconsistent snapshot projection; complete current causality is not established.
Consequence
Operators cannot distinguish absent evidence from operational health.
Repair pattern
Keep snapshot semantics and complete rendering consistent across stream transitions.
Regression check
Exercise Guide during updates; verify attention counts, all sections and missing-snapshot semantics.
Next acceptance
Reproduce the interaction at the exact deployed console revision.

BEV-0005 · Dated acceptance report · indexed 2026-10-09

Inspect supporting sourceRelated skill: Executable Operating Documentation

BUG-0006 · Google document retrieval · Tool transport

Expanded document JSON exceeds the response budget

Historical repair reported

Evidence scope: Recovered October read recovery; limits depend on tool and representation.

Inspect mechanism, repair and regression check
Trigger
A large document expands into a verbose serialised tool response.
Symptom
Read fails at the response-size boundary although the document is accessible.
Root cause / qualification
An unnecessarily expanded representation exceeds the response budget.
Consequence
Context recovery stalls or the document is falsely reported missing.
Repair pattern
Use compact native reads or bounded windows with explicit coverage.
Regression check
Read a representative large document compactly and verify intended content and bounds.
Next acceptance
Validate the present connector without assuming the old limit is universal.

BEV-0006 · Dated acceptance report · indexed 2026-10-09

Inspect supporting sourceRelated skill: Problem-Centred Institutional Memory

BUG-0007 · Tail Wagging · UI and contracts

First client theme differs from server HTML

Implementation reviewed

Evidence scope: Application-specific repair and reported browser checks; other providers need separate acceptance.

Inspect mechanism, repair and regression check
Trigger
A dark-theme route hydrates server light-theme markup.
Symptom
React reports a hydration mismatch or replaces content.
Root cause / qualification
Provider initial state differs between server and first client render; delays do not fix the contract.
Consequence
Rendering is unreliable and suppressed warnings conceal the cause.
Repair pattern
Match server and first-client theme state; apply browser preferences after hydration.
Regression check
Check light/dark mobile and desktop first renders and unexpected mismatch warnings.
Next acceptance
Run deployed-route hydration checks across supported browser conditions.

BEV-0007 · Public implementation and dated report · indexed 2026-10-09

Inspect supporting sourceRelated skill: React Hydration Contract Repair

BUG-0008 · DoneState · Authority and lifecycle

App metadata substitutes for customer permission

Implementation reviewed

Evidence scope: Merged/deployed repair recorded; current customer role is read and no fresh write objective established.

Inspect mechanism, repair and regression check
Trigger
Installation-token metadata decides customer write admission.
Symptom
Valid work is rejected or App capability stands in for customer writer authority.
Root cause / qualification
Installation capability and authenticated customer permission are treated as one principal.
Consequence
Objective admission is incorrect or exceeds customer authority.
Repair pattern
Check identity-bound customer permission separately from selected App capability.
Regression check
Test reader/writer customers, unselected repositories and missing App permissions independently.
Next acceptance
Complete a fresh permitted customer journey.

BEV-0008 · Public implementation and project state · indexed 2026-10-09

Inspect supporting sourceRelated skill: Dual-Principal GitHub App Admission

BUG-0009 · DoneState · Authority and lifecycle

Admitted writes recreate state after account deletion

Implementation reviewed

Evidence scope: Synthetic write-race acceptance is separate from real account cleanup and external revocation.

Inspect mechanism, repair and regression check
Trigger
A request admitted before deletion completes after cleanup.
Symptom
Registry or credential state reappears.
Root cause / qualification
Session rejection does not fence already-admitted durable writes.
Consequence
Erasure is incomplete under concurrency.
Repair pattern
Fence stale writes at registry and credential destinations using deletion/generation state.
Regression check
Pause admitted work, erase, resume, and verify no indexed state or credential resurrection.
Next acceptance
Revalidate indexed destinations and each actual deletion; track external revocation separately.

BEV-0009 · Public scoped acceptance evidence · indexed 2026-10-09

Inspect supporting sourceRelated skill: Account Erasure with Write Fencing

BUG-0010 · Tail Wagging · Durability

Duplicate-payment early return loses outcome evidence

Implementation reviewed

Evidence scope: Source contains reconciliation; complete production fault injection remains separate.

Inspect mechanism, repair and regression check
Trigger
Fulfilment succeeds, evidence fails, then the valid event is redelivered.
Symptom
A duplicate is acknowledged without repairing the missing outcome.
Root cause / qualification
Fulfilment idempotency short-circuits independently incomplete evidence.
Consequence
Successful paid work remains absent from the outcome ledger.
Repair pattern
Reconcile evidence independently of already-completed fulfilment.
Regression check
Fail evidence once, replay, and verify one fulfilment plus one recovered durable outcome.
Next acceptance
Inject controlled evidence failure and read unchanged customer effect counts.

BEV-0010 · Public implementation and dated audit · indexed 2026-10-09

Inspect supporting sourceRelated skill: Durable Business Outcome Reconciliation

BUG-0011 · Post-Once · Durability

Cursor advances before uncorrelated evidence is durable

Preventive failure mode

Evidence scope: Implementation protects this failure mode; not every production crash path was exercised.

Inspect mechanism, repair and regression check
Trigger
An outcome precedes publication evidence and holding persistence fails or fills.
Symptom
The source event can be skipped after cursor acknowledgement.
Root cause / qualification
Acknowledgement precedes successful durable holding.
Consequence
Delayed attribution is permanently lost.
Repair pattern
Persist validated holding before advancing the cursor; retain progress on failure.
Regression check
Inject full storage and persistence failure; verify cursor remains before the event, then recover fairly.
Next acceptance
Run current crash, capacity and restart acceptance with cursor readback.

BEV-0011 · Public implementation review · indexed 2026-10-09

Inspect supporting sourceRelated skill: Durable Business Outcome Reconciliation

BUG-0012 · Tail Wagging · Authority and lifecycle

Payment signature becomes campaign provenance

Implementation reviewed

Evidence scope: Signed attribution is present; descriptive correlation does not establish causation.

Inspect mechanism, repair and regression check
Trigger
A valid payment contains caller-supplied campaign labels.
Symptom
Unsigned labels enter attribution as authenticated source identity.
Root cause / qualification
Payment authenticity is incorrectly extended to unsigned provenance fields.
Consequence
Spoofed attribution corrupts business reporting and learning.
Repair pattern
Validate signed or source-owned attribution separately from payment evidence.
Regression check
Valid payment with unsigned labels stays paid but unattributed; trusted provenance can correlate independently.
Next acceptance
Check forged, absent and valid attribution at the deployed boundary.

BEV-0012 · Public implementation and dated audit · indexed 2026-10-09

Inspect supporting sourceRelated skill: Durable Business Outcome Reconciliation

BUG-0013 · One Click · UI and contracts

Omitted services produce schema-valid numeric filler

Candidate fix open

Evidence scope: PR 8 open at reviewed head dd72e2d; patch is not deployment or clean-account acceptance.

Inspect mechanism, repair and regression check
Trigger
A minimal brief omits services or supplies an empty list.
Symptom
The handoff passes schema with meaningless numeric service filler.
Root cause / qualification
Optional-input fallback preserves shape without useful output semantics.
Consequence
A successful response cannot support the intended website handoff.
Repair pattern
Handle omitted and empty inputs deliberately; assert meaningful output content.
Regression check
Check both omitted and empty services through the real public journey.
Next acceptance
Recheck merge/deployment and run fresh omitted/empty-input customer acceptance.

BEV-0013 · Public candidate patch · indexed 2026-10-09

Inspect supporting sourceRelated skill: Staged MCP Plugin Activation

BUG-0014 · OpsTruth · UI and contracts

External readers claim a closed-world boundary

Implementation reviewed

Evidence scope: Recorded merge/deployment and scans; provider publication and clean-account acceptance remain separate.

Inspect mechanism, repair and regression check
Trigger
A tool reads GitHub entities while claiming closed-world annotations.
Symptom
Activation metadata misdescribes actual behaviour.
Root cause / qualification
Read-only is confused with no external-world interaction.
Consequence
Provider activation and agent interpretation rely on inaccurate metadata.
Repair pattern
Annotate external readers accurately and retain compatibility checks for local validators.
Regression check
Inspect actual tools/list; separate external readers and local validators; retain version checks.
Next acceptance
Verify production discovery, conversational routing and settled publication separately.

BEV-0014 · Public merged repair · indexed 2026-10-09

Inspect supporting sourceRelated skill: Staged MCP Plugin Activation

BUG-0015 · OpsTruth · Measurement

Analytics helper mismatches SQL and sampling semantics

Implementation reviewed

Evidence scope: Helper repair established; historical telemetry includes owner/test traffic without user/cause markers.

Inspect mechanism, repair and regression check
Trigger
Telemetry is queried through legacy Workers Analytics Engine SQL.
Symptom
Request/parsing errors or unweighted aggregates misstate outcomes.
Root cause / qualification
Request format, response parsing and explicit legacy sample weighting are mismatched.
Consequence
Transport success or volume becomes a false usage or demand claim.
Repair pattern
Use actual raw-SQL and response contracts, weighting and transport/tool outcome separation.
Regression check
Validate real response, weighted aggregates and HTTP-success/tool-failure separation.
Next acceptance
Revalidate endpoint semantics and report sampling and identity limits.

BEV-0015 · Public merged repair · indexed 2026-10-09

Inspect supporting sourceRelated skill: Consent-Aware Plugin Measurement

BUG-0016 · Post-Once console · UI and contracts

Runbook warnings and dashboard fields drift

Implementation reviewed

Evidence scope: Contracts and saved acceptance reviewed; console fixture speed is not provider throughput.

Inspect mechanism, repair and regression check
Trigger
Docs, warnings and projections are maintained independently.
Symptom
Unknown evidence appears healthy or automatic work appears manual.
Root cause / qualification
Units, ownership and recovery semantics lack a shared executable contract.
Consequence
Operators take incorrect actions or accept unsupported health.
Repair pattern
Generate references/examples from versioned canonical field and warning contracts.
Regression check
Check warning semantics, units, missing values and projection parity against fixtures.
Next acceptance
Check deployed UI, generated docs and contract version together.

BEV-0016 · Public implementation and dated report · indexed 2026-10-09

Inspect supporting sourceRelated skill: Executable Operating Documentation

BUG-0017 · OpsTruth provider submission · Tool transport

Stale provider list contradicts settled submission detail

Cause unresolved

Evidence scope: Settled detail resolved the conflict; no duplicate retry occurred and root cause is unproven.

Inspect mechanism, repair and regression check
Trigger
The list still says not submitted after submission of the same draft.
Symptom
List and settled detail disagree about in-review state.
Root cause / qualification
State propagation is a candidate explanation; provider internals were not established.
Consequence
An agent may retry an already completed consequential action.
Repair pattern
Read same-draft settled detail and action receipt before deciding on a retry.
Regression check
Stale list plus settled in-review detail must trigger readback rather than duplicate submission.
Next acceptance
Retain exact draft/action locators and verify settled state before retry.

BEV-0017 · Dated observed state reconciliation · indexed 2026-10-09

Inspect supporting sourceRelated skill: State-Transition Evidence Contract

BUG-0018 · Cross-project · Evidence

Agent declares completion from a lower evidence layer

Preventive failure mode

Evidence scope: Cross-project continuation pattern; not a newly reproduced defect or proof every agent made the error.

Inspect mechanism, repair and regression check
Trigger
The highest result is a test, HTTP success, owner demo or merge.
Symptom
Agent reports deployed or customer-complete without matching evidence.
Root cause / qualification
Implementation, transport, release, publication and customer outcome become one done state.
Consequence
Later work resumes from an unsupported assumption.
Repair pattern
Record highest evidenced transition with revision, environment, principal and limits.
Regression check
Merged without deployment stays merged; owner success without fresh customer journey stays owner-scoped.
Next acceptance
Collect the missing deployment, provider or customer evidence for the claimed outcome.

BEV-0018 · Reviewed preventive pattern · indexed 2026-10-09

Inspect supporting sourceRelated skill: State-Transition Evidence Contract

BUG-0019 · Agent Shop · UI and contracts

Unsupported API write falls through to HTML success

Implementation reviewed

Evidence scope: CI run 37901366579 observed POST returning 200. The repair then passed all twelve method/surface checks on the local Node preview on 9 October. Production deployment remains unverified.

Inspect mechanism, repair and regression check
Trigger
An agent sends POST, PUT, PATCH or DELETE to a GET-only discovery endpoint.
Symptom
The framework returns HTML with HTTP 200 instead of rejecting the unsupported method.
Root cause / qualification
A missing method handler falls through to the application renderer; declaring GET alone does not enforce HTTP method rejection.
Consequence
Agent clients can mistake a successful HTTP response for an accepted write or handoff.
Repair pattern
Declare explicit write-method handlers returning 405 with Allow: GET on each read-only JSON route.
Regression check
Check POST, PUT, PATCH and DELETE against registry, integration brief and coding bug JSON: all twelve pairs must return 405 and Allow: GET.
Next acceptance
Verify method rejection on the exact deployed origin after release.

BEV-0019 · Public implementation and dated report · indexed 2026-10-09

Inspect supporting sourceRelated skill: Machine-Discovery Surface Engineering

BUG-0020 · Agent Shop · UI and contracts

Long metadata expands catalogue beyond phone width

Implementation reviewed

Evidence scope: CI run 37902732121 observed shop overflow at 390 pixels. The grid/card sizing repair passed full browser acceptance in run 37903191137 at head 79b1d784. Production deployment remains unverified.

Inspect mechanism, repair and regression check
Trigger
The catalogue renders long source paths inside nested grids at a 390-pixel phone viewport.
Symptom
The shop has horizontal page overflow despite metadata text using truncation.
Root cause / qualification
Grid tracks and cards retain automatic minimum content sizing; long metadata can impose its intrinsic width before truncation applies.
Consequence
Agent operators cannot inspect the complete catalogue within the phone viewport.
Repair pattern
Use explicit single-column mobile tracks and minmax(0,1fr), and allow catalogue content and cards to shrink with min-w-0.
Regression check
At phone width, document scrollWidth must not exceed window innerWidth on shop, controls and bug routes; preserve desktop grids and source-path truncation.
Next acceptance
Verify phone-width catalogue pages against the exact deployed release.

BEV-0020 · Public implementation and dated report · indexed 2026-10-09

Inspect supporting sourceRelated skill: Live Dashboard Update Verification