Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
When resuming an issue from preserved todos/context compaction, first rehydrate live GitHub issue + repo/artifact state before continuing. Treat restored todos as hints, not proof. If an expected plan/artifact path is missing, pivot to issue comments and repo search as source of truth. See references/resume-after-compaction.md.
This is the canonical execution route for approved GitHub issue work.
Use it only after planning is complete and the issue is approved.
Route summary
Canonical path:
Issue -> Plan -> User approves -> TDD implementation -> adversarial review -> commit/push -> comment/close issue
Canonical execution method:
Every non-trivial work item must be classified before execution as one of:
single-lane — small, tightly coupled, decision-heavy, or shared-file work handled in the main session
parallel-readonly — resource intelligence, planning evidence, review, validation, or risk analysis can run concurrently; the orchestrator owns synthesis and durable artifacts
parallel-worktree — implementation is already plan-approved and the write surface splits cleanly into isolated worktrees with explicit owned/read-only/forbidden paths
Use docs/standards/PARALLEL_FIRST_EXECUTION.md as the controlling standard. Parallelization is the default question for non-trivial work, not an afterthought; gates still control when implementation may begin.
Speed-optimization rule: when the user says Hermes issue execution is slow or asks for parallel agent optimization, do not answer only with generic delegation advice. Patch or consult the governing execution/planning surfaces so the optimization applies to every future issue: classify all non-trivial work upfront, parallelize read-only intelligence/review by default, and reserve write-capable lanes for plan-approved isolated worktrees with orchestrator-owned verification.
GitHub posting rule
As execution progresses, post meaningful GitHub updates at each major step.
Do not leave the issue silent until closeout.
Minimum posting cadence during execution:
execution start / scope confirmation
noteworthy findings from strengthened resource intelligence or pre-checks
if the issue is already done / invalid, post evidence before close
if new future issues are created, post the linkage immediately
after implementation + validation, post the landed change summary
before close, post verification evidence
Keep updates concise and cumulative. Prefer fewer structured updates over noisy chatter.
Future-issue capture rule
During execution, if you discover additional work that should not be silently absorbed into the current issue, capture it as a future GitHub issue.
Create it now when the split is clear, or mark it as a candidate when the orchestrator must decide timing or batching.
Claude agent-team prompt packaging rule
When execution is best handled by multiple agents, multiple terminals, or licensed/external machines, use Claude to package the work into a self-contained prompt pack for agent team(s).
Use this when:
work splits cleanly into non-overlapping streams
separate agents need explicit write boundaries
execution is overnight or unattended
a machine without Hermes context must execute the work
Prompt packs should be:
self-contained
explicit about scope boundaries
explicit about owned, read-only, and forbidden paths
explicit about tests, validation, and GitHub check-in expectations
explicit about GitHub authority limits
explicit about commit/push or comment-only behavior
explicit about future-issue capture expectations
Preferred supporting skills for this packaging:
overnight-parallel-agent-prompts
licensed-machine-prompt-orchestration
Parallel-first execution matrix
Choose the execution mode only after the entry gate and already-done pre-check are complete.
For non-trivial work, explicitly decide single-lane, parallel-readonly, or parallel-worktree; do not silently default to serial execution.
Do not delegate just because the issue is large; delegate only when ownership, validation, and GitHub reporting are all operationally clear.
Situation
Mode
Rule
single narrow fix, shared files, short feedback loop
single-lane
keep execution in the main session
uncertain implementation surface or active design decisions
issue may already be satisfied and only needs verification/closure
single-lane or parallel-readonly
keep final close decision in the main session and post evidence directly
multiple independent files/areas with clear acceptance slices
parallel-worktree
split into non-overlapping agent-owned streams after plan approval
overnight/unattended work with explicit validation steps
parallel-worktree for approved implementation; parallel-readonly for planning/review
package prompts so each stream can complete safely without live supervision
shared files, migration choreography, or likely merge contention
single-lane or serialized parallel-worktree
assign one writer for shared surfaces; do not run concurrent writes
wide but mostly read-only recon feeding one final implementation
parallel-readonly
delegate recon, keep final code integration central
approval/authz ambiguity, unclear write access, or missing repo/runtime prerequisites
single-lane blocker path
stop delegation and resolve authorization/readiness before any worker starts
Default to parallel-readonly for non-trivial recon/review/validation when it can reduce wall-clock time. Default to parallel-worktree for approved implementation only when ownership boundaries are crisp. Default to single-lane when boundaries are unclear or orchestration overhead exceeds the work.
Operational checkpoint before execution:
Confirm the issue is approved and in execution scope
Finish the already-done pre-check
Choose exactly one mode: single-lane, parallel-readonly, or parallel-worktree
Post the mode decision to GitHub when it changes who will execute or how check-ins will happen
Only then start coding or prompt packaging
Use this decision flow:
Choose single-lane when the next best step is design resolution, user decision, a tight single-threaded fix, or shared-file work with no safe serialization plan
Choose parallel-readonly when lanes can inspect, validate, review, or classify independently but should not write durable repo state
Choose parallel-worktree only when each approved implementation stream has explicit owned paths, explicit validators, explicit GitHub reporting ownership, and no likely write overlap
If you cannot prove one of those paths safely, stop and continue centrally or return for user/orchestrator decision
Required evidence for a delegated decision:
exact issue(s) or sub-slices assigned
owned/read-only/forbidden paths per stream
validator/test command per stream
orchestrator vs worker GitHub authority split
merge/integration point and final central verification owner
If that evidence cannot be stated compactly and unambiguously, do not delegate.
Delegated issue batch formation rules
When forming a delegated batch:
batch only issues/subtasks that can be proven non-overlapping in write surface
group by shared validator/test command only when that reduces risk rather than hiding scope
keep each batch small enough that one prompt can state exact paths, commands, and closeout expectations
isolate any migration, rename, or shared fixture work into a single owner
if two streams would touch the same file, same generated artifact, or same lock/state surface, keep them central or serialize them
Batch package must state:
issue numbers included in that batch
exact goal of the batch
owned paths for writes
read-only paths allowed for context
forbidden paths that must not change
validation commands that define done
Delegated prompt path contract
Every delegated prompt must include a path contract with these fields:
Owned paths: the only files/directories the delegated team may modify
Read-only paths: files/directories they may inspect for context but not edit
Forbidden paths: files/directories they must not modify under any circumstance without returning control
Rules:
owned paths must be explicit and minimal
anything not listed as owned is read-only by default unless marked forbidden
shared config, lockfiles, generated outputs, release metadata, and issue-tracking artifacts should be forbidden unless one batch is explicitly assigned ownership
if delegated execution discovers a required change outside owned paths, stop, report it, and convert it to a future issue or orchestrator decision
GitHub authority split
Default authority split during execution:
orchestrator owns issue start notes, scope changes, blocker posts, future-issue creation/linking, final landed summary, and issue close
delegated teams own execution evidence inside their assigned stream packet and should return checkpoints to the orchestrator unless the prompt explicitly authorizes direct GitHub posting
if direct delegated posting is authorized, limit it to assigned issue progress updates and require the orchestrator to post the final integrated closeout
Do not let multiple delegated teams post overlapping status narratives on the same issue without an orchestrator merge point.
Decision checkpoint rule
At the end of each major execution step, explicitly choose one:
continue current issue
create future issue
stop for user decision
stop for blocker
Do not carry ambiguity across steps.
Create a future issue when the discovered work is:
outside the approved plan or acceptance criteria
a separate bug, hardening task, cleanup, migration, or risk remediation
blocked by another dependency or decision
important enough that it should be tracked rather than left in a comment
When creating a future issue:
use a precise title and scope
describe why it was split from the current issue
link back to the current issue
mention the new issue number in the current issue comments and final closeout note
When a delegated team surfaces future work, return it in a compact handoff block:
Future issue candidate:
Why split: <scope/risk/dependency reason>
Evidence: <file/test/behavior proving it>
Suggested owner/path:
Blocks current issue?: yes/no
The orchestrator decides whether to create the GitHub issue immediately or queue it for the next batch.
Blocker protocol
If execution is blocked, post a GitHub update immediately with:
blocker summary
impact on scope, delivery, or verification readiness
missing dependency/decision/input
whether a future issue was created
whether execution is paused or rerouted
If the blocker is substantial and not resolvable inside current scope, create a future issue or dependency-tracking issue instead of letting the current issue absorb it silently.
Runtime / machine-availability blocker rule:
When an approved issue explicitly requires a licensed tool, external machine, or runtime that is not available in the current session, do not fake implementation and do not treat repo-side helper code as satisfying the issue.
First, gather local evidence that sharpens the blocker:
verify the required runtime/tool is actually unavailable on this machine
verify any relevant repo-side precursor code/tests that already exist
confirm whether the acceptance target still requires the missing machine/runtime for final proof
Then post a GitHub update that clearly separates:
what was verified locally
what remains blocked on the missing runtime/machine
why the issue is not yet satisfiable from the current environment
If another approved issue is executable from the current environment, pivot immediately to that next bounded target rather than stalling.
Typical examples: OrcaFlex/OrcaWave/ANSYS issues requiring licensed-win-1, CFD/OpenFOAM issues requiring absent solver binaries, or hardware-dependent validation unavailable from the current host.
Entry condition
Only start execution when the issue is already labeled status:plan-approved.
If the issue is still unplanned or only in status:plan-review, route back to gh-work-planning first.
Treat execution entry as a hard gate, not a soft preference.
Before any code change, branch creation, or delegation package, confirm all of the following:
approval/authz: the issue has explicit plan approval and you are authorized to execute in this repo/worktree/environment
scope: the approved plan and acceptance target are identifiable enough to verify against
workspace readiness: the intended repo/worktree, branch context, and required policy files are known
validation readiness: you know the primary test/validation path or know that closure will rely on deterministic inspection evidence
GitHub readiness: you can post the required execution/check-in/closure updates, or the orchestrator for doing so is explicitly defined
Entry decisions:
if approval is missing, stop and route back to planning/approval
if authorization or execution environment is not safe/available, stop and post a blocker or hand back to the orchestrator
if validation cannot yet be identified, stay in central recon until a verification path exists
if the issue appears already satisfied, do not begin implementation; follow the already-done pre-check to a verification-first closeout decision
GitHub check-in at entry:
when execution is allowed to proceed, post a concise start note with scope boundary, execution mode if already known, and immediate validation intent
when execution cannot proceed, post the stop reason and next route (planning, blocker, or orchestrator decision)
Local plan-marker gate in worktrees
Local plan-marker gate in any checkout
In workspace-hub-style repos, a GitHub status:plan-approved label may still be insufficient for implementation if local hooks enforce .planning/plan-approved/ markers.
Before launching implementation in the current checkout (main checkout or fresh worktree), verify all of the following in that same checkout:
.planning/plan-approved/<issue>.md exists locally
the marker text uses neutral/operator approval wording (not Worker session, auto-approved, or self-approved)
the marker is committed in that checkout before write-capable Claude/Codex execution begins
Safe sequence for approved issue execution:
ensure you are in the exact checkout that will perform the writes
if .planning/plan-approved/<issue>.md is missing locally, create it there with neutral approval wording tied to the real user approval signal
commit the marker locally in that same checkout
only then launch implementation
Worktree example:
create the worktree from main
write .planning/plan-approved/<issue>.md inside the worktree
commit the marker locally in that worktree
only then launch Claude/Codex for implementation
Main-checkout example:
even if the issue is already status:plan-approved on GitHub and the user explicitly says to continue, create/commit the local .planning/plan-approved/<issue>.md in the active checkout before implementation if it is missing there
Why this matters:
the plan-approval hook evaluates the local checkout state, not just GitHub labels
a marker created only in another checkout, or created but not committed yet, may still be treated as missing or self-approved
committing the marker first avoids the implementation session getting blocked mid-run by the local plan gate
committing the marker first avoids Claude getting blocked mid-run by the local plan gate
Practical pre-commit check before your first implementation commit:
confirm .planning/plan-approved/<issue>.md exists
confirm it is newer than .planning/STATE.md in the current checkout
if not, refresh or recreate the marker with the same user-approved content before staging implementation files
then retry the commit
Applies to
a single approved GitHub issue
a batch of approved GitHub issues
Hard constraints
Read repo policy files first
Respect the approved plan
Strengthen resource intelligence before changing code when uncertainty remains
Use TDD for code/script changes
Validate before commit
Use adversarial review for non-trivial changes
Commit/push with verification evidence
Close the issue only when landed or proven invalid/already satisfied
Pre-check: is the issue already done?
Before writing code, run a verification-first closure check.
Do not treat "looks done" as enough; require evidence that would justify closing the issue right now.
Minimum pre-check sequence:
inspect the expected deliverable surface: files, flags, config, docs, behavior, or output named by the issue/plan
if this is a verification-first closeout, first make sure the local checkout actually reflects the candidate landed state (git fetch, compare HEAD vs origin/main, and fast-forward/rebase or switch to a clean worktree as needed). Do not conclude a file/test is missing from the repo until you have ruled out local checkout staleness.
run the most relevant targeted tests/validators if they exist
read issue comments/linked PRs for prior implementation or landing evidence
inspect recent commits/history when comments or repo state suggest the work may already have landed partially or fully
compare current repo behavior against the approved acceptance target, not just against changed files
make an explicit decision: already done, not done, or uncertain
Evidence required to conclude already done:
at least one direct proof of the deliverable in the current repo state
at least one verification artifact: passing test output, deterministic inspection result, or runtime/CLI evidence
enough linkage to explain why this issue is satisfied now: commit hash, merged PR, landed file path, or issue comment reference when available
acceptance-criteria coverage sufficient to justify closure, not just partial progress
for stale-open issues closed from a verification-first path, prefer two git proofs when possible:
history proof: commit containment / ancestry on origin/main
content proof: current target artifact matches the intended landed content (not just history topology)
Practical stale-open closeout rule:
git merge-base --is-ancestor <commit> origin/main is a stronger landed check than broad branch-listing alone.
Commit containment is not always enough in repos with merge races or follow-up reverts; add a content-parity check on the primary deliverable when feasible (for example git diff <commit> origin/main -- <path> expecting no relevant delta, or deterministic inspection of the current file contents).
For design/doc issues, the narrowest validator may be the document's own normative contract applied to the landed artifact rather than a test runner.
Prefer posting the closeout comment first and then closing the issue in a separate step. This avoids the gh issue close --comment race/drop hazard if another actor closes the issue before your close call completes.
Decision outcomes:
already done -> post a GitHub update with the evidence bundle, state that implementation is already satisfied, and close the issue from the verification-first path
not done -> post a concise note only if the pre-check revealed meaningful findings, then continue to execution mode selection
uncertain -> do not start broad implementation yet; stay in central recon until you can prove already done or not done
Important verification-first nuances:
If the issue asked for an intermediate tranche (for example: tests-only, xfail-only, or a narrowly scoped sub-step) but the repo has already moved past that point with stronger landed work, treat that as already done when the stronger landed state still satisfies the requested acceptance target. Say explicitly that the repo exceeded the originally proposed intermediate stop.
For parent / claim-boundary / umbrella issues, open follow-on hardening issues do not automatically block closeout. If the parent acceptance criteria are satisfied by landed child work and the remaining items are clearly future hardening rather than missing parent requirements, close the parent with explicit linkage to the still-open follow-ons.
GitHub posting for this stage:
if the issue is already done, the comment must include what was checked, what passed or was observed, and what artifact proves the work landed
if the pre-check changes implementation understanding materially, post that finding before coding or delegation
if uncertainty remains after reasonable checks, post the uncertainty and next verification step instead of pretending execution has started cleanly
This avoids wasting an execution cycle on already-complete work and prevents premature closure without proof.
Strengthened resource intelligence during execution
Execution is not blind implementation. Re-open intelligence work when needed.
Strengthen resource intelligence whenever you hit uncertainty about:
where the real implementation surface is
whether a validator/runtime/docs mismatch already exists elsewhere
whether the approved plan missed an adjacent dependency
whether another issue/PR/comment already solved part of the problem
whether an upstream/external dependency constrains the fix
whether a required source artifact exists only in an off-repo, licensed, confidential, or local corpus location
Resource intelligence during execution may include:
broader code/test/doc search
GitHub issue/PR/comment history review
recent commit inspection
session recall
upstream or package/API verification
searching for alternate or legacy terminology
verifying controlled off-repo sources and safe tracked derivatives without copying restricted raw artifacts into the repo
Post a short GitHub note when this changes the implementation understanding materially.
Licensed/off-repo source resolution
When a blocker is resolved by finding a licensed or off-repo source artifact:
verify the source with deterministic evidence: absolute path or controlled route, title/metadata when available, parser/report counts, and generated-artifact sanity checks
post a GitHub decision-ledger update that separates blocker resolution, source location, safe tracked derivatives, raw artifacts that must not be committed, and interpretation limits
patch any stale plan text that still says the source is unresolved
rerun focused plan review only if the source update changes scope, acceptance criteria, or validation path
require fail-closed implementation behavior if the source/provenance/citation sidecar is unavailable
Do not overstate generic/reference source data as asset-specific or client-specific. Keep raw licensed workbooks/PDFs/client files off-repo; track only safe derived artifacts and routing documentation.
For licensed standards workbooks used by generated engineering reports, also use references/generic-ocimf-workbook-adapter-tdd.md: it covers adapter-first TDD, no-corpus-leakage boundaries, generic/reference limitation language, and cross-surface HTML/DOCX/PDF verification.
Execution pattern per issue
1. Post execution start
Comment that execution has started for the approved issue only after the entry gate, already-done pre-check, and execution-mode decision are complete.
Include:
scope boundary for this execution pass
chosen mode: central, delegated, or hybrid
who owns GitHub check-ins if workers are delegated
immediate validation target or first failing test to be used
2. TDD-first loop
Do not start implementation until there is a failing proof for the requirement or a documented reason why deterministic inspection is the only valid path.
Loop:
choose the smallest test or validator that can fail for the approved requirement
write or update that test first
run it in targeted mode and capture the failure evidence
if it passes unexpectedly, stop and choose one:
already done -> return to the verification-first closure path
wrong test -> tighten the test until it proves the real gap
wrong scope -> stop and re-check the approved plan before coding
make the minimal change that should turn the test green
re-run the same targeted validator immediately
repeat until the requirement turns green without widening scope
Evidence required before leaving the loop:
test/validator name or command
failing result was observed at least once unless deterministic inspection is the approved proof path
final targeted pass result after the fix
Stop/continue rule:
continue only while each loop iteration reduces the known gap to the approved acceptance target
stop when the next change would require unapproved scope, touching forbidden paths, or guessing about missing acceptance details
For scripts/config/docs:
use the narrowest executable or static assertion available
if no executable test exists, require deterministic inspection evidence and name exactly what artifact proves the requirement
3. Minimal implementation pass
Implement the smallest change that satisfies the approved plan and the failing proof.
Rules:
prefer one requirement slice at a time over broad rewrites
keep unrelated cleanup out unless it is required for correctness or safety
if adjacent work is necessary, state why it is inseparable from the approved fix
if adjacent work is useful but not required, create a future issue instead of expanding scope
4. Targeted validation gate
Run validation in escalating order and stop at the first failing gate:
fast local correctness checks relevant to the changed surface
syntax/type/lint/unit target directly affected by the change
requirement proof
the exact test(s)/validator(s) used in the TDD loop
safe behavior confirmation when applicable
dry run, CLI invocation, deterministic output check, or manual inspection with explicit expected result
broader regression check only when risk justifies it
nearby test file, package-level suite, or other bounded regression pass
Validation discipline:
start with the smallest command that can reject the change
broaden only when the risk surface justifies it
do not substitute a broad green suite for missing targeted proof
if a validator flakes or is environment-blocked, do not wave it through; either stabilize it, document the blocker, or stop
for composite readiness/checker work, separate component acceptance from umbrella host readiness: a repo-placement checker can be dispatchable: true while the broader readiness command is fail because the active worktree is dirty/ahead/behind during implementation; record both results and do not misclassify component acceptance as failed unless the broader failure is part of the issue acceptance criteria
when hardening checkers with injected/synthetic state, distinguish an absent key from an explicit bad value: treat {"upstream": None} as missing-upstream evidence only when the key is present, not when a partial injected state omits the field for unrelated tests
Required evidence bundle:
exact commands run
whether each command was targeted or broader regression coverage
pass/fail result and any deterministic inspection artifact
explicit note for anything intentionally not run and why
Acceptance-criteria traceability
Before closeout, map each acceptance criterion to explicit proof:
acceptance criterion -> test name(s)
acceptance criterion -> validation command(s)
acceptance criterion -> deterministic inspection evidence when tests are not sufficient
If any criterion lacks proof, the issue is not ready to close.
Post a GitHub progress update after validation when the issue spans meaningful implementation work.
Multi-agent verification gate
When work was delegated across agents or terminals, do not close on per-agent success alone.
Before commit/closeout, the orchestrator must run a unified verification gate:
confirm each delegated stream stayed inside owned paths
review diffs for hidden overlap, missing integration, or conflicting assumptions
run the agreed validation commands from the main session/environment
map acceptance criteria across the combined result, not just per-stream outputs
decide whether any cross-stream discovery must become a future issue before closeout
Required integration evidence:
per-stream completion status: done, partial, blocked, or rejected
per-stream validator result and any path-contract deviation
final integrated validation output from the orchestrator session
explicit statement of whether new future issues were created from cross-stream findings
If unified verification is incomplete, keep the issue open and post the remaining gap.
Do not commit/push/close on worker-local evidence alone.
5. Adversarial review gate
For non-trivial changes, run a review pass focused on:
correctness vs issue acceptance criteria
hidden behavior mismatches
shell safety / state handling / regression risk
adequacy of tests
whether discovered extra work should become future issues instead of expanding scope
artifact/output contract drift across every emitted surface, not just the primary code path
For report, audit, inventory, or generated-artifact work, the adversarial review must challenge JSON summaries, Markdown sections, helper scripts, closeout reports, fixture/dry-run modes, and untrusted/partial-state behavior. If a safety condition suppresses authoritative findings, verify summaries and generated artifacts are suppressed or clearly marked too; do not accept a green test suite that only covers the primary findings list while secondary outputs can still mislead closeout.
For generated engineering/scientific artifacts, add a model-to-artifact consistency challenge: do not rely on string-presence tests or report claims alone. Verify the numerical/physical model used in analytical summaries is the same model emitted into solver decks, CSV/JSON profiles, plots, and reports. Add at least one regression assertion at an interface/boundary/representative point that would fail if the artifact uses a simplified or stale formula (for example cumulative layer resistance vs linear-thickness interpolation, load magnitude in a solver card vs reported power, or boundary-condition value vs report text).
Classify the result explicitly:
PASS -> no material objections remain
MINOR -> optional improvements only; can continue if documented
MAJOR -> correctness, safety, scope, or validation gap remains; must fix before commit/push
Required review evidence:
what was challenged
what failed or almost failed under adversarial reasoning
what was changed or consciously accepted as non-blocking
Stop/continue rule:
if review returns MAJOR, return to the TDD/validation loop
if review returns MINOR, record the residual risk level, residual risk, or future-issue candidate
continue only on PASS or documented MINOR
Post a GitHub update if review materially changed the solution.
6. Commit/push gate
Do not commit or push just because the code looks done.
Commit is allowed only when all are true:
the TDD proof is green
targeted validation evidence is recorded
acceptance-criteria traceability is complete
adversarial review is PASS or documented MINOR
artifact hygiene check is complete
for delegated/hybrid work, the orchestrator integration gate is complete
Before push, verify:
the commit message is issue-linked and accurately scoped
no unrelated files are staged
the pushed state matches the validated state
if repo policy requires PR flow, route to that flow instead of silent direct push
Push behavior:
push immediately after a completed issue or tightly related approved pair
if push fails, resolve the Git problem without weakening the validation standard
if the validated tree changes after commit but before push, re-run the affected validators before pushing
Recovery / rollback rule
If validation or review fails after implementation:
do not push a weak or unverified fix
revise the change and re-run validation, or
revert/reset the local change if the approach is wrong, then post a GitHub update describing the recovery path
7. Unified closeout + issue comment
Post one final structured GitHub closeout comment before closing the issue.
Change summary: concise description of what changed or what was proven
Acceptance criteria: criterion -> proof mapping
Validation: exact commands/results and deterministic inspection evidence
Adversarial review: PASS / MINOR and any residual risk
Residual risk level: Low / Medium / High
Git evidence: commit hash(es), push status, and PR link if applicable
Future issues: created numbers or none
Residual risks: explicit list or none
GitHub posting rule for closeout:
if landed, post after validation/review and after the change is committed/pushed or otherwise landed per repo policy
if already satisfied or invalid, post the proof bundle used to justify closure
if blocked-rerouted, post why the issue remains open or what successor issue now tracks the work
Close only when one of these is true:
the fix is landed with evidence
the issue is already satisfied with evidence
the issue is invalid/obsolete with evidence
Keep the issue open when:
push/landing is incomplete
acceptance traceability has a gap
validation or review evidence is missing
a blocker was identified without a resolved reroute path
After close/push, verify and repair status labels explicitly. Auto-closing through a Closes #NNNN commit can leave stale workflow labels such as status:plan-approved on a closed issue. Read the final issue labels, remove conflicting status:* labels, and apply the terminal label used by the repo (for workspace-hub, usually status:done) before reporting final gate state.
For generated inventory/report closeouts, preserve authoritative input snapshots rather than overwriting them with a simplified helper fixture. If a helper needs planning-time paths, parse the existing snapshot schema (for example paths.filesystem_only_active) and add tests for that schema. Treat accidental snapshot shrinkage as artifact drift: restore the snapshot, update the helper adapter, rerun targeted tests, regenerate only the intended closeout report, and then commit.
8. Multi-agent closeout/integration reporting
When delegated or hybrid execution was used, append a compact integration block to the final closeout comment:
Execution mode: delegated or hybrid
Streams: list each stream/batch and status: done / partial / blocked / rejected
Path contract check: confirm compliance or list deviations
Integration verification: commands/results run by orchestrator in the final environment
Cross-stream findings: future issues created or none
Closeout owner: confirm orchestrator posted the final summary and performed the close/no-close decision
If any stream is partial, blocked, or rejected, do not present the issue as complete.
State exactly what remains and whether it stays on this issue or moved to a future issue.
Useful batching heuristic
Combine issues only when they truly share the same code path or tests.
Do not combine unrelated issues just to reduce commit count.
For broader waves:
finish the active issue cleanly
use read-only recon for next candidates if needed
decide whether to implement centrally or package as Claude prompts for agent team(s)
if packaged to agent teams, enforce zero git contention with explicit owned/read-only/forbidden path boundaries
close false positives directly when evidence supports it
create future issues instead of letting discoveries disappear between waves
Execution patterns worth keeping
Large test coverage issues
Use research first, then parallel implementation only when many files are involved and fixture/context is clear.
Always finish with unified verification in the main session.
Library integration issues
Use a compact pattern:
check installed/version
install if needed
smoke test the API
add evaluation script + tests
verify
commit
close issue with specifics
create follow-up issues for deferred hardening or broader adoption work
Inventory/report prerequisite slices in nested repos
For approved umbrella/decomposition issues where the first executable slice is a durable inventory or evidence artifact, do not jump directly to source remediation.
Use this pattern:
Confirm the parent issue is status:plan-approved and the local approval marker exists in the executing checkout.
Post an execution-start comment that explicitly limits the first pass to the inventory/report artifact and says source remediation remains in child issues.
In the nested target repo, run the exact failing command from the approved plan and capture raw output to a transient path (for example /tmp/...). Treat a non-zero exit as expected when inventorying known debt.
Generate a checked-in report under the nested repo with: exact command provenance, exit code, parsed finding count, grouped rule-family counts, dominant outlier classification, non-outlier counts, representative findings, and an explicit note that /tmp raw output is transient/non-durable.
Add deterministic assertions for the report content (parent issue link, child owner, exact command, total/outlier counts, and cleanup guidance) before committing.
Commit and push the nested-repo artifact separately from workspace-hub governance/review artifacts; verify both nested HEAD and origin/main match.
Run a compact adversarial implementation review on the report slice. If reviewers find metadata/table issues, patch the report, commit/push a follow-up, and record the review artifacts in workspace-hub.
Post progress comments to both the parent umbrella and the relevant child issue, making clear which prerequisite is complete and what remains open.
Common pitfalls:
Table grouping code may accidentally mix file paths and directory areas; define module-area tables as directory paths only.
UTC regeneration timestamps can cross the local session date; distinguish inventory capture date from report regeneration timestamp.
Do not close the parent umbrella after a prerequisite inventory lands; keep closure tied to the child issues and final proof gate.
Verify-and-close issues
Some issues are best solved by verification rather than code changes.
If evidence shows the requested state already exists, comment with proof and close.
Approved umbrella/decomposition issues
Some approved issues are umbrella/decomposition parents, not direct source-remediation tickets. When the approved plan explicitly splits implementation into child issues:
Do not close the parent just because the first prerequisite lands.
Pick the smallest approved parent-owned slice (often durable inventory, coordination artifacts, or proof scaffolding) before child source remediation.
If the implementation artifact lives in a nested repo, commit/push it in that nested repo and record review/coordination artifacts in the parent orchestration repo separately.
Post progress to both the parent umbrella and the relevant child issue so ownership stays visible.
Keep source remediation inside the child issue boundaries; do not mix a pathological outlier child with a safe-rule cleanup child.
After landing a prerequisite artifact, run adversarial implementation review on that artifact. If reviewers return MINOR metadata/reporting findings, fix them in a follow-up commit before calling the prerequisite complete.
Parent closeout remains blocked until every child stream and the final proof owner satisfy the approved acceptance criteria.
For lint-inventory prerequisites, a reusable TDD/documentation pattern is:
generate a durable inventory report with exact command provenance, exit code, parsed finding counts, outlier classification, non-outlier counts, transient /tmp warning, and cleanup guidance
add a small test that asserts the report preserves the key provenance/decomposition strings
run the exact inventory command and the focused report test
record implementation-review artifacts in the orchestration repo if the report itself lands in a nested repo
Interruption and tool-budget recovery
If execution is interrupted by context compression, tool-call exhaustion, provider limits, or an enforced final-response boundary while validation or closeout is still pending:
Do not present the issue as complete or imply closeout succeeded.
Preserve the active gate state explicitly: implementation complete vs validation in progress vs commit/push/close pending.
Name the exact next checkpoint to resume from, starting with git status, changed-file inspection, targeted validators, adversarial review, then commit/push/issue closeout.
When the user or runtime explicitly says no more tools are allowed, do not attempt more inspection. Produce a compact restart checkpoint from already-observed evidence only: worktree path, branch, issue state/labels if known, dirty files, untracked files, implemented surfaces observed in diffs, what has not been verified, and exact next commands/checkpoints for the next session.
Separate observed evidence from inferred state. If tests, review, commit, push, or closeout were not run in the current visible context, say so plainly and keep the issue open in the status narrative.
If skill loading fails due to ambiguous duplicate skill names, do not retry the same skill_view arguments. Continue with already-loaded governing skills or load the repo-local/canonical path explicitly in the next session.
Treat the resulting final answer as a restart checkpoint, not a substitute for GitHub closeout evidence.
When the runtime reports preserved active todos after context compression or max tool-call exhaustion, fold those todos into the checkpoint but do not upgrade them to evidence. State the live gate still pending (for example validation in progress, adversarial review pending, commit/push pending), the exact worktree/branch if observed, dirty files if observed, commands already run with results, and the next tool-backed resume sequence. Do not claim issue closure, review completion, or push success unless those were observed in the current visible context.
For interrupted generated engineering/report closeouts, use references/interrupted-engineering-report-closeout.md before resuming commit/push/close. It adds the artifact-surface and source-truthfulness checks that are easy to miss after a focused green test run.
Git pitfalls
When an approved parent issue has become an umbrella/decomposition contract and child issues own the source-level remediation:
Do not drift into child source edits from the parent pass, even if the failing command is easy to inspect.
Run the parent pre-check against the live nested/target repo state: clean main, exact approved command, current counts, and presence/absence of the durable inventory artifact.
If the parent deliverable is an inventory/provenance artifact, use TDD against that artifact: write a failing test that locks the required command provenance, expected-red baseline, grouped counts, outlier classification, child issue ownership, and transient-vs-durable evidence warning; then make the minimal doc/report change to pass.
Re-run the exact failing command after the doc/test change to prove source-remediation scope was not silently absorbed and the baseline remains the expected child-owned red state.
Post the parent update without closing the umbrella unless all child completion/green-gate acceptance criteria are actually satisfied; also post a child-boundary note when the artifact belongs to a child issue.
If unrelated local dirt blocks a clean-main start, preserve it with narrow temporary stashes rather than doing branch cleanup or mixing it into the issue execution.
Git pitfalls
If push fails due to remote mismatch after a valid local commit:
first verify whether the remote actually advanced despite the error: git fetch origin <branch> && git rev-parse HEAD origin/<branch>
if HEAD == origin/<branch>, treat the push as landed and do not retry/pull unnecessarily
if GitHub rejects with cannot lock ref ... is at <new-sha> but expected <old-sha>, that can still mean the push landed while another process updated the local remote-tracking ref; immediately fetch and compare HEAD vs origin/<branch> before retrying
if the remote is genuinely ahead or divergent, inspect the new remote commits, then git pull --rebase only when it is safe for the current scope, and push again
If unrelated edits block the push, stash them temporarily when safe. If stash silently fails or branch switching is blocked by unrelated dirt, use a clean temporary worktree from the target branch to make docs-only/closeout commits; record that the original checkout remains dirty so the next operator does not assume a clean main state.
Artifact hygiene
Before commit, check for runtime artifacts and temporary outputs.
Do not stage them unless the issue explicitly requires them.
When generating Markdown closeout/review artifacts from shell, avoid unquoted heredocs that contain backticks, command-substitution syntax, or command-like text. Use write_file, a single-quoted heredoc delimiter (<<'EOF'), or escape every substitution surface; otherwise Bash may execute inline artifact prose and corrupt evidence files before commit.
For issues that implement in one sibling repo but publish generated deliverables into another checkout or embedded output tree, use references/mixed-repo-output-artifact-closeout.md. Key rule: classify the actual Git boundary for every output path, commit implementation and packaged outputs separately when they live in different repos, stage only explicit paths from dirty orchestration repos, and parse/extract DOCX/PDF content before claiming client-ready artifacts.
Recommended progress report
After each wave or checkpoint, report:
issues completed
new work vs verified-already-done
commit hashes pushed
validation/review gate status
future issues created
what remains feasible next
whether anything is only local vs already pushed
running totals when helpful
Preferred companion skills
github-issues for issue comment/close/create actions
github-pr-workflow only when the repo or user explicitly wants PR-first execution
Legacy compatibility
Older guidance may reference workspace-hub-batch-issue-execution.
Treat that as a deprecated alias/reference for this route, not the primary route name.
