Bounded swarm executor for kb-plan output. Runs all ready vertical slices from the dependency DAG, parallelizing only independent/isolated work, with TDD enforcement, scope gates, and HITL pauses. Use when the user says 'kb work', 'work the plan', 'execute the plan', 'run the KB pipeline', 'execute all slices', or wants guided execution of a planned feature.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Bounded swarm executor for kb-plan output. Runs all ready vertical slices from the dependency DAG, parallelizing only independent/isolated work, with TDD enforcement, scope gates, and HITL pauses. Use when the user says 'kb work', 'work the plan', 'execute the plan', 'run the KB pipeline', 'execute all slices', or wants guided execution of a planned feature.
argument-hint
[path to KB manifest, or blank to find latest]
KB Work - Bounded Swarm Slice Executor
Run all vertical slices from a kb-plan manifest by pulling the safe ready set
from the dependency DAG. Keep each slice tied to its acceptance criteria,
enforce the requested verification mode, and pause on HITL tasks.
Quick Start
Read the KB manifest.
Validate the dependency DAG and statuses.
Verify the manifest gate_ledger allows kb-work; repair or block if plan-to-work is not passed.
Confirm execution once unless the user already asked to run/execute/work the manifest.
Execute the safe ready set without asking between slices.
Update the manifest after each slice so the workflow is resumable.
After all runnable slices are terminal, write the work-to-complete gate and immediately invoke kb-finalize <manifest-path> unless the user explicitly said to stop before finalization.
Input
#$ARGUMENTS
If input is empty: Scan docs/plans/ for the most recent *-kb-*-manifest.md file. If found, use it. If no manifest exists, scan todo.md, docs/brainstorms/, and docs/requirements/ for one active unplanned source. If exactly one exists, invoke kb-plan <source> with execution intent, then return to kb-work <manifest-path>. If none or multiple plausible sources exist, ask for the feature/source to slice; do not execute unplanned work.
If input is a path: Read the manifest at that path.
If input is a handoff: Do not execute the handoff directly. If it links a docs/plans/*-kb-*-manifest.md, use that manifest. If it contains only phases, workstreams, bullets, or broad next steps, stop and invoke kb-plan to create vertical slices first.
If input is a feature description, broad task, or "go straight to work" request instead of a manifest path: invoke kb-plan <input> with execution intent, then return to kb-work <manifest-path>. kb-work only executes KB manifests with per-slice plans and an initial expected_files forecast.
Continuous Completion Loop
When the user invokes kb-work with execution intent, this skill owns the loop
until the work is truly terminal.
Terminal means one of:
every slice is done or intentionally skipped, then kb-complete has run
through review, follow-up resolution, proof, memory, and cleanup;
the only remaining slices are blocked, human-required, or parked, with
exact resume criteria recorded in todo.md and the manifest;
the user explicitly says to pause or stop.
Default WIP is the safe ready set, not one slice. A slice is ready when its
blockers are done or skipped, its status is pending, and it is not marked
as a serial-only gate. Dispatch ready slices together only when the runtime gives
each active slice an isolated checkout/context or an equivalent write-isolation
guarantee. On a shared checkout, WIP is one mutating slice at a time.
expected_files is a forecast, not proof of disjointness. If active slices
observe or claim writes to the same path, serialize or requeue one slice before
continuing. Observed overlap beats planned disjointness.
Do not stop at weaker milestones:
"the current slice passed";
"all slices are done";
"tests passed";
"review started";
"I wrote the summary."
Those are progress states. The next action is still to continue the loop, either
to the next runnable slice or to kb-complete.
If a repo has a project-specific done.md contract such as "can't stop til its
done", treat it as this same terminal rule. Do not create a new done.md from
the global skill; use todo.md, todo-done.md, manifests, and handoffs as the
KB state system unless the repo already opted into done.md.
Pre-Flight
Read the manifest - parse the YAML frontmatter to get the ordered slice list.
Validate DAG - confirm no cycles in blockers, all referenced slice IDs exist, and all slice files exist.
Validate gate ledger - read gate_ledger. The plan-to-work gate must be passed and its allowed_next_action must name this manifest. Run kb-gate/scripts/check_gate_ledger.py <manifest-path> --gate plan-to-work --allowed-next "kb-work <manifest-path>" before execution. If the gate is absent, pending, blocked, stale, or the checker fails, stop and route to kb-plan/kb-gate to repair it before execution. Do not execute from a manifest that lacks a passing gate.
Validate objective contract - when the manifest opts into objective_contract: true, it must have a top-level done_check, and every runnable slice must have proof_check or a valid no_check_reason. Do not require a durable model_route or attempt_tier: new plans record planned correction/authority difficulty and risk, while attempt selection happens immediately before dispatch from live evidence. Legacy model_route values may remain readable as hints only. When cmd/kbcheck exists, run go run ./cmd/kbcheck manifest-contract --manifest <manifest-path> before execution. Otherwise verify those fields and terminal slice gates directly, record manifest-validator: unavailable, and block on any missing/false proof_check, absent done_check, or unproved terminal status. The portable skills do not require the Go maintainer harness. If validation fails, route back to kb-plan/kb-gate; do not continue from a manifest that can self-report completion without objective proof.
Validate slice contracts - each slice plan must have expected_files, verification, blockers, status, and acceptance criteria. New slice plans should also have test_level, functional_risk, model_tier, and proof_check or no_check_reason. New plans do not contain model names, route aliases, source preferences, adapters, endpoints, or transports. Treat legacy tiny or model_route fields as compatibility hints only; do not require them on new plans and do not freeze a provider/model in durable plan state. If core fields are missing, stop and route to kb-plan; do not infer a manifest from a phase list. If only test_level or functional_risk is missing on an older plan, invoke kb-functional-test to classify them before execution.
Load the context packet - for a non-trivial slice, read its packet before broad repo search or delegation. When cmd/kbcheck exists, validate JSON packets with go run ./cmd/kbcheck context-packet --packet <packet.json>. Otherwise verify the required fields directly and record packet-validator: unavailable; the portable skills do not require the Go maintainer harness. If required source, constraint, proof, search-policy, or escalation data is missing, route back to kb-plan instead of making a cheap worker rediscover the repo. Legacy tiny/mechanical slices may use the plan itself when it records why no packet is needed.
Check status - skip any slices already marked done. Resume from the first safe ready set.
Check worktree - note dirty or untracked files before executing so unrelated user changes are not staged or reverted.
Read optional execution policy - next-lower AMR attempts are disabled by default while the pilot is unpromoted. Enable them only for an explicit pilot/opt-in or amr.lower_tier_attempts: enabled; otherwise start at the planned tier. Read any personal project source preference from user-local kb-models state; an unsaved preference means automatic. Ordinary work never pauses for a routing-priority question. Offer and persist automatic, self-hosted-first, or native-first only during explicit kb-map setup or kb-models requests. Do not collect connection details here.
Read active landmines — if docs/context/landmines.md exists, read only Active Landmines and carry any relevant failure modes into slice execution and verification. If a slice touches an owner_surface, treat that landmine as a hard guardrail until the slice proves the verification condition or explicitly leaves it active.
Sync with board — read todo.md and confirm its status table matches the manifest. If they diverge, the board wins — another agent may have updated it. Reconcile the manifest from the board before proceeding.
Confirm once only when needed: If the user did not explicitly ask to run/execute/work the manifest, ask: "Ready to execute N remaining slices in order. Proceed?" If the user already asked to execute, continue without this prompt.
After initial execution starts, do not ask before moving from one safe ready set
to the next.
Treat statuses as:
Status
Action
pending
Eligible once blockers are done or skipped
done
Skip
blocked
Stop and ask whether to retry, skip, or abort
human-required
Waiting on human action; continue unrelated runnable slices if possible
parked
Intentionally out of bounds today; only a human promotes back to active
skipped
Skip but keep visible in the summary
Board Sync Protocol
todo.md is the live execution board. Update it at every status transition:
Event
Board Update
Starting a slice
Set status to 🔧 in_progress
Slice completes
Set status to ✅ done
Slice blocked
Set status to 🔒 blocked + reason in notes
Slice needs human action
Set status to 🛑 human-required + exact ask
Slice parked by human
Move to 🧊 Parked / Cold Storage with reason
Slice skipped
Set status to ⊘ skipped
All slices done
Prepend compact summary to todo-done.md, then remove completed feature section and routine completion logs from todo.md
Active handoff files under docs/handoffs/active/ are restart packets. Create or update one whenever work stops, blocks, or changes the next recommended action. Move completed handoffs to docs/handoffs/done/.
Multi-agent rules:
Before claiming a slice, re-read todo.md. If another agent set it to 🔧, do not claim it.
The board is the source of truth — not chat history, not the manifest. If the board says done, it's done.
Update the board BEFORE starting work (claim) and AFTER completing work (release). This prevents two agents from working the same slice.
Also update the manifest to stay in sync, but if they conflict, the board wins.
Do not use root Work Log as a permanent archive. During execution, add notes only when they help a later agent resume: blockers, verification commands, durable memory impacts, or non-obvious decisions. Routine "slice complete" and verification-success notes belong in todo-done.md at feature completion, not in todo.md.
Blocked is not parked. Use 🔒 blocked for dependencies, another-agent waits, tool failures, or missing inputs. Use 🧊 Parked / Cold Storage only for work a human intentionally deferred out of scope.
Run-State Events
If the manifest or goal ledger points to .kb/runs/<goal-slug>/, append a row
to .kb/runs/<goal-slug>/route-history.jsonl when kb-work starts, completes,
blocks, or requeues a slice.
Use state_changed: true or progress_key only after a manifest status, gate,
proof artifact, or board pointer actually changed. Before choosing another KB
lane for the same run, validate the history with:
go run ./cmd/kbcheck run-state --history .kb/runs/<goal-slug>/route-history.jsonl
If the guard fails, stop routing and repair the loop by refreshing context,
re-planning the unit, or asking the smallest human question.
Live Route Selection
Routing is a work-time decision, not a plan rewrite.
For each run/ready slice, use the executable sequence:
Resolve the router once per session without crawling the filesystem: use
kbrouter/kbrouter.exe when command -v/Get-Command succeeds; otherwise
use $HOME/.kb/bin/kbrouter on POSIX or $HOME\.kb\bin\kbrouter.exe on
Windows when present and executable. If neither exists, record
router-unavailable: binary-not-found and use the planned-tier current-model
path when policy permits. Never auto-download, auto-build, or search sibling
repos/drives. A custom installer --router-dir must be placed on PATH.
Use the resolved executable for every command below.
For a routed decision, call kbrouter dispatch with the selected primary
alias, optional next ordered alias, and unique direct-child packet/output/
receipt/handoff names containing the slice ID.
If discovery/select is missing, incompatible, or returns unavailable, record
router-unavailable: <reason>. Use the returned degraded-current decision or
the current model only when policy permits; require pauses that slice instead.
Pass --attempt-tier only after the current master establishes Step 2.6
eligibility; the selector validates candidates but does not infer task
suitability.
Discover or reuse exactly one run-scoped catalog before the first routed
slice. Reuse it while the host/configuration fingerprint is unchanged;
refresh it only when that fingerprint changes.
Never ask a worker-tier, model-by-model, or source-priority question during
ordinary work. Merge what the active host can select with eligible user-local
extras; an unsaved source preference is automatic.
Treat model_tier as the planned correction/authority tier. New tiers are
small, medium, and large; planner is an orchestration/review role.
Legacy tiny maps to the small lane as a compatibility hint only.
Before dispatch, decide whether this slice is eligible for one explicit
next-lower-tier attempt. Eligibility requires settled intent, bounded files,
interfaces and authority, an objective proof that can reject a bad result,
low destination/trust risk, and exact escalation triggers. Code or a file
extension alone is never eligibility.
If eligible and attempts are enabled, select the attempt route from live
evidence at the next lower tier. If not, begin at the planned tier. Within
either tier, try another qualified same-tier route before moving to the
planned tier, a higher tier, or the current model as policy permits. The
attempt is explicit runtime policy, not an inferred fallback or a durable
plan field.
Apply the saved user-local project source priority only after eligibility. automatic
lets the current master choose by evidence, self-hosted-first prefers
eligible user-local extra routes, and native-first prefers eligible
host-native routes. Preference never overrides trust, authority, tools,
context, risk, or proof and never hard-pins a route.
A higher or planner-grade model may execute lower-tier work when it is
independently eligible or explicitly requested by the user.
Security, auth, data-boundary, or process-boundary work must not be silently
down-routed or sent to a less trusted destination.
use <model> prefers that route first when it satisfies the slice's bounded
eligibility, trust, and authority, then keeps safe fallback. It overrides the
saved user-local project source preference for this run.
require <model> bypasses automatic attempts and pauses only that slice if
the exact route is unavailable. It is the only hard pin.
prefer self-hosted (prefer local shorthand) or prefer native is a
source preference inside trust and risk constraints.
ignore model routing bypasses attempts and uses the current model with
ordinary proof gates only.
A generic or named subagent spawn without an exact selector cannot claim the
requested model. Record actual route/model/session only from dispatcher or
host evidence; otherwise provenance is unknown or unavailable.
Existing correct work with missing or mismatched provenance is not blocked or
redone for telemetry. Investigate once, record what is knowable, then let
independent proof govern completion.
Show a compact preview only when routed slices exist. For a lower-tier attempt,
show at most one plain line, for example
Trying Small for a bounded, objectively proved change; Medium correction fallback. Otherwise group by
the planned live route and bounded fallback path, for example
Medium — <preferred-route> -> <same-tier-route> -> current(degraded): 004, 007.
This preview is dispatch intent for the current ready set only. If nothing is
routed, say nothing about models.
Ready-Set Ordering
Execute by repeatedly pulling the safe ready set from the dependency DAG:
Build a map of slice_id -> slice.
For each pending slice, check all blockers.
The candidate ready set is every pending slice whose blockers are complete.
Exclude serial-only slices from co-dispatch when other ready slices exist:
can_continue_other_slices: false, HITL-critical gates, destructive
approvals, browser/e2e contention without isolated sessions, and any slice
with an active write lease collision.
Dispatch the remaining safe ready set in isolated contexts when available.
If no isolation is available, run the same ready set one mutating slice at a
time while preserving the ready-set order.
If pending slices remain but none are runnable, mark the manifest blocked and
report the dependency problem.
Execution Loop
For each slice in dependency order:
Continuous Execution Rule
Ready sets should run continuously once execution has started.
Do not ask "Proceed to execute slice-N?" between slices. Move to the next
safe ready set automatically after:
slice status is updated;
board and manifest are synced;
required deterministic checks pass;
QA/repair gates pass or are not applicable.
Pause only when a real gate requires it:
HITL decision or missing value that cannot be generated safely;
blocked/human-required/parked slice with no unrelated runnable work;
destructive command approval;
out-of-scope file edit or diff-scope failure;
QA/repair exhaustion or stuck loop;
dependency deadlock;
observed write overlap that cannot be serialized or requeued safely;
user explicitly asked to pause or stop.
Step 1: Check HITL Flag
If hitl: true:
Present the slice title, description, and the specific question/decision needed.
Classify the HITL item before stopping:
critical-path — later slices depend on this decision/access/input.
parallel-blocker — this slice is blocked, but unrelated slices can continue.
final-validation — human judgment is useful before release, but not needed for development.
agent-runnable-with-inputs — human only needs to provide values; the agent can run the check.
Stop only the dependent path. If unrelated slices are runnable, mark this slice blocked or human-required, update todo.md and the manifest, then continue those slices.
When marking a slice human-required, parked, or blocked, persist: owner, blocked_reason, resume_when, next_agent_action, human_action, can_continue_other_slices, and parked_at.
Record the user's decision in the slice plan.
Update manifest status to done for this slice only if the decision completes the slice.
Continue to the next runnable slice.
Missing test inputs are not a reason to ask the user to manually test. If test_inputs are missing:
Ask for the specific missing value.
Use safe generated or fixture values when acceptable.
If the input blocks only this slice, mark this slice human-required or blocked and continue unrelated runnable slices.
Resume the slice after the value is available and run the verification yourself.
Step 2: Deepen If Thin
If the slice plan has fewer than 3 acceptance criteria or no test scenarios:
Run a lightweight deepening pass on this single slice.
Add concrete test scenarios and likely file paths.
Keep the pass bounded; do not re-plan the whole feature.
Step 2.5: Test-Level Classification
Before editing, ensure the slice has a recorded test obligation:
test_level: none, unit, integration, functional-api, functional-cli, functional-browser, or full
functional_risk: none, narrow, broad, or full
model_tier: small, medium, or large (tiny remains readable only as a legacy hint that maps to small)
If test_level or functional_risk is missing, stale, or contradicted by the
acceptance criteria or expected_files, invoke kb-functional-test with the
slice plan. Record the result in the slice frontmatter or notes before
implementation. If model_tier is missing in a manifest that opted into
model_tier_contract, route back to kb-plan to repair the slice contract.
Use smaller worker routes for this classification only when the active host can
actually select them. Keep the task bounded: classify the slice, audit existing
tests for mocked theater, and suggest the narrowest deterministic proof.
Escalate to the main model for complex architecture/auth/security/flaky async
decisions or repeated test failures.
Model-tier boundaries:
Tier
Correction authority fits
Move higher when
small
narrow mechanical edits and straightforward tests with clear acceptance criteria
hard decomposition, architecture, broad debugging, final synthesis/review
proof is impossible or human-only input is required
The tier does not change the proof bar. It records the authority and capability
required to correct or take over the slice if an initial attempt fails. It is
not a claim that every first attempt must run at that tier.
Record actual runtime/model, turns, input/output/cache tokens, proof result, and
packet sufficiency when the host exposes them. Missing host telemetry is
unavailable, not zero. Raw values remain authoritative; weighted cost scores
must be versioned and reported beside proof outcomes.
Step 2.6: Adaptive Model Routing (AMR)
AMR owns one proof-triggered attempt and correction loop:
planned correction tier + bounded packet + objective proof
-> eligible? try one next-lower tier : begin at planned tier
-> deterministic proof passes? preserve result
-> proof fails? prepare bounded correction handoff; ordinary planned-tier execution
-> isolated correction runner unavailable? never dispatch into the live checkout
Immediately before a ready slice runs:
Use the current master to decide attempt eligibility from the packet and
policy. The selector must not infer suitability from “code,” a file suffix,
or model price.
If eligible for the explicit pilot, pass runtime attempt_tier to kbrouter
and select only a dispatch-qualified next-lower-tier route that satisfies tools, context, trust,
destination, and risk. Otherwise begin at model_tier. Apply the saved
user-local project source preference only among eligible live routes; plans provide no
model or transport hint.
Run the slice's narrowest deterministic proof. Proof is the validator; the
planned-tier model is the correction authority, not the validator.
If proof passes, keep the attempt without a stronger-model rewrite and
continue ordinary QA, regression, review, and completion gates.
If proof fails, prepare a compact ordinary-fallback surgical handoff summary
for the planned-tier-or-higher authority containing:
accepted_result:"<what already satisfies the contract>"failed_criterion:"<one observed failure>"failure_location:"<file + symbol/line/hunk when known>"allowed_change:"<smallest behavior/files/hunks>"preserve_invariants: ["<interfaces, passing behavior, tests, user decisions>"]
relevant_interfaces: ["<callers/callees/schema/DOM/API>"]
proof:command:"<exact command>"exit:"<status>"artifact:"<path/hash>"failure:"<bounded excerpt>"current_diff:"<compact diff or artifact reference>"attempt_ledger:"<routes, receipts, results, no credentials>"corrective_diff_only:trueproof_to_rerun: ["<focused proof>", "<regression proof>"]
This summary is an agent-facing handoff contract, not the typed
internal/modelrouting.CorrectionPacket, and it must never be serialized as
that packet or passed to dispatch correction. The typed envelope is a
future executable boundary that additionally requires a failed-proof routing
receipt, driver-owned authority hashes, bounded artifact references, and an
independent hunk-local oracle. Until a host-owned isolated workspace, proof
runner, and compare-and-swap apply path exist, automatic correction dispatch
must fail closed and the current driver performs separate ordinary
planned-tier execution from the original scope. Do not claim preserved-work
savings.
A future isolated correction model may edit only the allowed behavior/hunks,
preserve only work with independent hunk-local acceptance, return the
corrective diff, and rerun focused plus regression proof. It must not rewrite
the whole file to improve style.
When no hunk-local oracle exists, the defect is unlocalizable, an authority or
interface boundary changes, or focused correction fails, abort the surgical
pilot and record separate ordinary planned-tier execution, re-plan, or
request HITL. Never silently restart or claim preserved-work savings.
An HTML implementation from an approved design with browser assertions may be
eligible. Choosing the design, interpreting philosophy/policy, or implementing
code without an adequate oracle is not. An explicit require pin or
ignore model routing bypasses automatic lower-tier attempts; use and
prefer self-hosted/prefer local/prefer native remain bounded by trust,
authority, and proof.
Record observable attempt/correction telemetry when available: planned tier,
user-local project source preference, attempt tier/route, correction route, receipts,
proof result, changed hunks, tokens/time, and escalation reason. Missing values
are unavailable. The resolved actual route belongs in the receipt. Routing
receipts remain telemetry; ordinary checks and protected oracles decide
acceptance.
Do not use unit just because it is cheaper. Use unit only when unit-level proof can fail for the user-facing bug or behavior. If a unit test could pass while the workflow is broken, require integration or functional proof.
Hard gate: when kb-functional-test auto-classifies a test level, the agent must not downgrade it. If expected_files includes .tsx, .jsx, .vue, or .svelte, or if non-UI files change behavior primarily reached through the rendered app UI, the slice is test_level: functional-browser.
When test_level is functional-browser, these steps are mandatory:
Start or connect to the running app.
Use Playwright to navigate to the actual feature route/screen in the rendered UI. Use CDP or the repo/platform authenticated browser transport only when Playwright cannot access an authenticated/corporate route.
Exercise the happy path with real clicks, keyboard input, form input, navigation, or other visible controls.
Capture screenshots of key pass/fail states and assert observable rendered outcomes after the action.
Clean up artifacts created during testing: test data, screenshots/traces when no longer needed, temp files, and browser state per repo QA cleanup rules.
Backend/API/unit checks may supplement this proof, but they cannot replace it. This gate cannot be skipped, overridden, or deferred.
Step 2.9: Regression Snapshot Gate
Before starting a new slice, invoke kb-regression-snapshot verify before Scope Lock and before editing implementation files.
Verify all previous snapshots under .kb/snapshots/.
If any previous snapshot fails, STOP before new slice execution.
Mark the current slice 🔒 blocked with the failing snapshot path, target, expected vs observed result, and artifact/log path.
Do not continue to implementation, QA, or the next slice until the regression is resolved, parked by the human, or explicitly skipped.
This gate catches entropy between slices. It cannot be skipped, overridden, or deferred.
Step 3.0: Scope Forecast and Ledger
Before executing the slice, load the declared scope forecast and keep a live ledger of actual files touched. expected_files guides the first pass; it is not a literal allowlist.
Read expected_files from the slice plan's frontmatter.
If expected_files is empty or missing, route back to kb-plan to repair the slice plan before execution. Do not execute from a phase list or raw task with no file forecast.
Expand the forecast with convention-matched test files. For each entry in expected_files, automatically include its corresponding test file based on project naming conventions:
Source file
Auto-allowed test file(s)
src/foo.py
tests/test_foo.py, test/test_foo.py
src/Foo.tsx
src/Foo.test.tsx, src/__tests__/Foo.tsx
lib/foo.rb
spec/foo_spec.rb, test/foo_test.rb
pkg/foo.go
pkg/foo_test.go
Test files that do not correspond to any expected_files entry may still be valid when current code or acceptance criteria require them; record them as discovered files.
Before opening any file for writing, classify it against the slice intent:
Finding
Action
File is listed in expected_files or is a convention-matched test
Proceed with the edit.
File is not listed, but current code shows it is directly required for this slice's acceptance criteria
Proceed, and add a manifest note: scope-discovery: <file> - <why required>.
File is generated by the repo's normal tooling, formatter, snapshot, lockfile, or test convention
Proceed, and add a manifest note: scope-discovery: <file> - generated/tooling.
File would change product scope, architecture direction, dependencies, migrations, auth/security boundaries, destructive behavior, or another slice's promised behavior
STOP for HITL or route back to kb-plan to amend the manifest before editing.
File is opportunistic cleanup or unrelated improvement
Do not edit. Park it in todo.md or a follow-up note.
Log the forecast in the manifest notes: scope-forecast: loaded N expected files + M convention-matched tests.
This gate pairs with Step 3.6 (Diff-Scope Verification). The point is traceability, not pretending the planner knew every file in advance.
Step 3: Execute
Use a fresh sub-agent when the platform supports delegated execution and the user has permitted it. Otherwise execute the slice locally while keeping the scope limited to this slice.
Immediately before dispatch, choose the route for this ready slice from the
live run catalog. Apply Step 2.6 once: use an explicitly eligible next-lower
attempt_tier, or begin at the planned model_tier. A failed attempt goes to
the planned tier as separate ordinary execution using the correction packet as
a bounded handoff. Do not dispatch automatic correction into the live checkout
or claim preserved-work savings. Use a higher tier or current model only after
the planned path cannot safely finish and policy permits.
Quoting sanity rule: when shell commands, file operations, or test assertions involve nested quotes, escaped quotes, or more than one quoting context, write the content to a temp file and execute/read that file instead of constructing the command inline.
If you are escaping an escape, you are doing it wrong. Write to a file, execute the file.
For multi-line JSON, SQL, HTML, scripts, or config blocks, use heredoc syntax or a temp file rather than inline quoting.
Do not build JSON strings inside shell commands inside assertion code. Write JSON to a temp file and read it.
Do not construct CSS selectors through mixed-quote string concatenation. Use template literals or parameterized locator helpers.
Use references/execution-prompt.md as the per-slice execution prompt/checklist. Load it only when starting a slice. When that slice runs Go inside a workspace sandbox, also load references/go-sandbox.md; its environment applies to Go shell invocations, never the agent launcher.
Step 3.1: Protected Oracle Gate
If the slice plan or manifest declares protected_oracles, enforce the
anti-cheat contract before implementation changes:
Identify every oracle file: tests, fixtures, scorers, snapshots, schemas, or
contracts that define expected behavior.
If the oracle is new or intentionally changed for this slice, create/update it
before implementation and prove RED when practical.
Record the oracle SHA256 in the slice plan or manifest after the oracle is
accepted.
After the SHA is recorded, do not edit that oracle unless the plan is
explicitly amended with a new reason and a new SHA.
Before marking the slice done, recompute oracle hashes. Any unexpected hash
change blocks the slice.
If protected_oracles is empty, continue with the declared verification mode.
Do not invent a protected oracle when expected behavior cannot be known before
implementation.
Step 3.5: System-Wide Test Check
Before marking a slice done, pause and ask these questions — vertical slices cut through all layers, so side-effects matter:
Question
What to do
What fires when this runs? Callbacks, middleware, observers, event handlers — trace two levels out from your change.
Read the actual code for callbacks on models you touch, middleware in the request chain, after_* hooks.
Do my tests exercise the real chain? If every dependency is mocked, the test proves logic in isolation — not interaction.
Write at least one integration test that uses real objects through the full callback/middleware chain.
Can failure leave orphaned state? If your code persists state before calling an external service, what happens when the service fails?
Trace the failure path. If state is created before the risky call, test that failure cleans up or that retry is idempotent.
What other interfaces expose this? Mixins, DSLs, alternative entry points.
Grep for the method/behavior in related classes. If parity is needed, add it now.
When to skip: Leaf-node changes with no callbacks, no state persistence, no parallel interfaces. Purely additive changes (new helper, new partial) take 10 seconds to confirm "nothing fires, skip."
Step 3.6: Diff-Scope Verification
After a slice completes, verify that the files actually changed are explainable by the slice's acceptance criteria. The agent does not self-report; the actual git diff is checked and recorded.
Get the actual diff:
git diff --name-only $(git merge-base HEAD main)..HEAD
This produces the list of files modified by this slice relative to the branch baseline.
Load the forecast scope from the slice plan's expected_files frontmatter field plus any scope-discovery: notes recorded during execution. Also load any protected_oracles and their recorded hashes.
Compare and enforce:
Apply the same convention-matched test file expansion as Step 3.0.
Finding
Action
Changed file is forecast, convention-matched, generated/tooling output, or recorded scope-discovery
Proceed.
Changed file is unforecast but directly required by the acceptance criteria and was not noticed before editing
Record scope-discovery: <file> - <why required> before proceeding.
Changed file expands product scope, architecture direction, dependencies, migrations, auth/security boundaries, destructive behavior, or another slice's promised behavior
STOP. Amend the manifest through kb-plan or get HITL before proceeding.
Changed file is unrelated cleanup or opportunistic improvement
Revert or park as follow-up before proceeding.
Forecast files were not changed
Treat as a completeness signal, not a failure. If the slice still satisfies acceptance criteria, record scope-forecast-unused: <file> - <why not needed>.
If a changed file is a protected oracle and its SHA changed after protection,
STOP unless the manifest or slice plan explicitly records an oracle update
reason and the new SHA.
Log results in the KB manifest under the slice's notes field:
If the slice plan has no expected_files field, route back to kb-plan to repair the plan before continuing. Do not execute a slice with no forecast at all.
This gate is mandatory. It cannot be skipped, overridden, or deferred, but it records justified discovery instead of blocking every unforecast file.
Step 3.7: Destructive Command Guard
Before executing any shell command during a slice, check it against this blocklist:
Blocked Pattern
Why
rm -rf, rm with recursive/force flags
Irreversible file deletion
git push --force / git push -f
Rewrites remote history
git reset --hard
Discards uncommitted work
DROP TABLE / DROP DATABASE / TRUNCATE
Irreversible data loss
git clean -fd
Deletes untracked files permanently
Bulk delete operations on files or data
Mass irreversible removal
Overwriting production config files
Environment-breaking changes
When a blocked command is detected:
STOP. Do not execute.
Show the user the exact command and explain why it's blocked.
Wait for explicit HITL approval before proceeding.
If running in autonomous mode (no HITL available), skip the command and log in the manifest notes: destructive-guard: blocked <command> — no HITL available
This is enforcement, not a warning. The agent MUST NOT execute destructive commands without explicit human confirmation. This gate cannot be skipped, overridden, or deferred.
Step 3.8: KB QA (all slices)
Invoke kb-qa with the current slice context. QA runs:
Lint check on forecast and discovered files for the slice
Browser verification against acceptance criteria for frontend slices and any backend/API/state slice whose changed behavior is reachable through the UI
If any check fails, kb-qa invokes kb-repair for surgical fixes (progress-based, 5-iteration cap). If repair exhausts or gets stuck, STOP — do not proceed to the next slice.
Truly backend-only slices skip browser checks but still run lint. Do not classify a slice as backend-only when the behavior being changed is primarily proven by using the app UI.
This gate is mandatory. It cannot be skipped or deferred.
Step 3.9: Figma Design Sync (UI slices only)
If the slice involves UI changes and Figma designs exist:
Implement components following design specs
Use the figma-design-sync agent iteratively to compare
Fix visual differences identified
Repeat until implementation matches design
Skip this step entirely for non-UI slices.
Step 4: Verify and Update
After the slice completes:
Check result
If yes: update manifest status: done for this slice and update the body table.
If no and repair/progress is still possible: run kb-repair or a bounded fix loop, then retry verification.
If no progress remains: update manifest status: blocked or parked, add failure notes and resume criteria, then continue unrelated runnable slices.
Before setting status: done, write or update a gate record
slice-<slice_id>-to-done. This gate must prove: implementation finished,
scope check passed, protected oracles were preserved or explicitly amended,
deterministic checks ran, functional/browser checks ran when required,
regression snapshot captured, memory impact was classified, and the slice's
proof_check passed or its no_check_reason was accepted by the manifest
contract. If any proof is missing, leave the slice blocked and set
allowed_next_action to the missing proof step.
For manifests with objective_contract: true, run
go run ./cmd/kbcheck manifest-contract --manifest <manifest-path> after
adding the slice gate and before changing the slice status to done. A
failing manifest contract blocks completion even when local tests pass.
Sync board — update todo.md status for this slice (done or blocked). Append validation note.
Run verification
Invoke kb-check for deterministic verification.
Prefer existing scripts, lint, typecheck, tests, browser checks, builds, and CI-equivalent commands over LLM inspection.
If a full suite is too expensive or unavailable, run the narrowest deterministic check that proves the slice and record why.
Invoke kb-functional-test whenever test_level is integration, functional-api, functional-cli, functional-browser, or full, or when user-visible/cross-boundary changes appear despite a lower test level.
If the slice fixed a known failure with a runnable sensor, record RED-before-GREEN proof with go run ./cmd/kbcheck accept --check <check.json> --trace .kb/trace.jsonl.
For UI-reachable changes, record UI proof: route/screen exercised, interaction performed, assertion made, browser transport used, and screenshot path when applicable. Do not mark the slice done with backend-only proof if a UI path exists.
After Step 3.8 QA passes, invoke kb-regression-snapshot capture <slice-id> with a compact spec for what changed. Store .kb/snapshots/<slice-id>.json.
Record test-level: <value>; functional-risk: <value>; proof: <command/artifact>; snapshot: <path/result> in the manifest notes.
Record the slice's proof_check command/artifact/result or accepted
no_check_reason in the manifest notes.
Assess memory impact
Classify the slice as memory-impact: none, operational, or durable.
none: cosmetic, copy, formatting, lint-only, or isolated tests with no behavior/architecture change.
operational: active state, blockers, verification commands, or handoff instructions changed. Update todo.md or the active handoff.
For durable changes, add a manifest note: memory-impact: durable; areas=<areas>; docs=<candidate docs>; refresh=pending.
If the affected doc is obvious and small, update it now. Otherwise leave refresh=pending for Step 5.
Optional commit
If the user asked for commits, stage only the manifest file for status updates and commit it separately.
Continue to the next runnable slice.
Step 5: Completion
When all slices are done or intentionally skipped:
Update manifest status: completed.
Run final verification.
Run kb-gate if verification, QA, repair, or functional-test checks surfaced P0/P1/P2/P3/P4 issues. P0/P1 block completion while unresolved, but safe/actionable blockers should be rectified before asking the user. P2/P3/P4 do not block by severity alone.
Write work-to-complete in the manifest gate_ledger. Required proof: every non-skipped slice has a passing slice-<id>-to-done gate, skipped slices have explicit reason, final verification command/result is recorded, no unresolved P0/P1 exists, board/manifest are synced, and scope-verified-files is populated. Set allowed_next_action: kb-finalize <manifest-path> and run kb-gate/scripts/check_gate_ledger.py <manifest-path> --gate work-to-complete --allowed-next "kb-finalize <manifest-path>". If the gate is not passed or the checker fails, do not invoke kb-finalize.
Refresh project memory — if any slice has memory-impact: durable or refresh=pending, run kb-map refresh before archiving. Update affected architecture, operation, decision, research, todo.md, and handoff pointers. Add manifest note: kb-map-refresh: done or kb-map-refresh: skipped - <reason>.
Archive to board — move the feature summary from todo.md to todo-done.md. Prepend at the top of the archive file with a completion date header.
Prune active board — remove the completed feature section from todo.md. Also remove routine work-log entries for the completed feature from todo.md; keep only still-active rows, 🔒 blocked rows, 🛑 human-required rows, the 🧊 Parked / Cold Storage section, or handoff-pointer items.
Report summary:
KB <name> — all slices complete.
- N slices executed
- S slices skipped
- M tests added
- K files changed
Verification: <command/result>
Next: kb-finalize runs automatically for review, documentation, and learning.
Persist scope context — collect the forecast and discovered file lists from each slice's notes field (the scope-check: and scope-discovery: entries from Step 3.6). Write the combined actual changed file list to the manifest frontmatter as scope-verified-files so kb-finalize can pass it to kb-review without re-deriving.
Post-work steps (kb-review, compound, learn, evolve, cleanup) are handled by kb-finalize. After all slices are done or intentionally skipped, invoke kb-finalize <manifest-path> automatically unless the user explicitly asked to stop after work execution. The kb-work run is not finalized until kb-finalize reaches its Done section or records a real blocker. It does not publish; user-facing kb-complete owns configured delivery.
Failure Handling
Situation
Action
Lower-tier AMR attempt fails proof
Prepare the compact surgical correction handoff, then run separate ordinary planned-tier execution; automatic live-checkout correction dispatch is disabled and preserved-work savings are not claimed
No hunk-local oracle, unlocalizable defect, boundary expansion, or failed focused correction
Abort the surgical pilot and record separate ordinary planned-tier execution, re-plan, or request HITL; do not claim preserved-work savings
Slice execution fails with progress possible
Run kb-repair or a bounded fix loop, then retry verification
Slice execution fails with no progress
Mark only that slice blocked/parked, write resume packet, continue unrelated runnable slices
Test suite fails after a slice
Run kb-repair; if stuck, mark affected slice blocked/parked and continue unrelated runnable slices
HITL critical-path pause
Present context, wait for user, record decision
HITL not on critical path
Park the slice and continue unrelated runnable slices
User says "abort"
Mark remaining slices as pending, stop
User says "skip"
Mark slice skipped, continue to next runnable slice
Resume Support
KB work is resumable:
Manifest tracks status per slice.
Re-running kb-work on the same manifest picks up where it left off.
Already done or skipped slices are not rerun.
Success Criteria
No slice runs before its blockers are complete.
Manifest frontmatter and body table reflect actual slice status.
Each completed slice has verification evidence recorded in the final response or failure notes.
No unrelated files are staged, committed, reverted, or overwritten.
Integration
Input from:kb-plan
Deepening: Use kb-research only when a slice has a material unresolved uncertainty before execution.
Execution engine: Fresh sub-agents when available, local execution otherwise
Verification: Preserves and reruns protected test oracles for tdd slices; load standalone tdd only for explicit test-first coaching.
Protected oracles: When declared by kb-plan, freezes behavior tests, fixtures, scorers, snapshots, or contracts before implementation so the target cannot move silently
Deterministic checks: Invokes kb-check before a slice is marked done
Functional checks: Invokes kb-functional-test for user-visible and cross-boundary behavior
Post-work finalization: Automatically invoke kb-finalize after all slices are done or intentionally skipped
