git:20260912.c5e08b8 to git:20261020.cf173e8

+ **Start with instruction reads.** After this entry point loads, your next tool calls read the three shared protocols below and the selected mode procedure. Do not batch them with repository searches or source/document reads. Reading only this SKILL.md does not open investigation. Complete the profile-loading checkpoint before investigation, tests or edits; only the declared bounded predicate routing below may inspect content earlier. A filename-only scope operation is not that checkpoint.

+

+ - **Always read change context** at [shared-change-context.md](shared-change-context.md), including when no profile matches. It defines the factual record used before editing and refreshed for independent review.

+ Both modes share the same execution core (Step 2 onward) — instruction loading, investigation and change context, implementation, verification, review.

+ - [ ] Implementation addresses intended outcomes, preserved contracts and hard task/delivery requirements

+ - [ ] Verification/tests pass for the claimed scope; unverified conditions are explicit

+ - [ ] Refreshed context and evidence supplied to review; findings resolved or permitted follow-up recorded durably

+ - [ ] Code completion distinguished from release/migration/activation conditions

+ By default, review checkpoints use **isolated mode** (`/kk:review-code:isolated`, `/kk:review-spec:isolated`). This is mandatory because the implementing session has authorship bias — the same model that wrote the code produces weaker reviews of it. Isolated mode spawns an independent sub-agent with no prior exposure to the implementation.

+ **Mandatory order — instructions before action.** The flow below is strictly sequential. Load this skill, the selected mode procedure, shared protocols and every resolved profile instruction before source/requirement investigation, edits (including task-status edits), tests or implementation decisions. Entry procedures collect minimal task/file metadata solely to select scope and profile guidance; they do not analyze requirement prose. After basic instructions load, the sole early-content exception is routing for a declared detection or conditional-load predicate: inspect at most approximately 16 KiB per candidate file, log the predicate/path, and conservatively load an instruction when its conditional is undecidable within that bound. Routing permits no behavioral analysis, full-diff read, findings, edits or tests.

+ **Phases:** load basic instructions and scope metadata → detect profiles and load their content → investigate requirements/source and establish change context → implement and verify → refresh evidence and review → assess completion. Detailed investigation has one entry point in Step 2; mode entry procedures do not duplicate it.

+

+ ### Step 1: Load Instructions and Mode Context

+ Read the shared protocols in Conventions, determine mode (see §Modes), then read the appropriate mode file and follow its entry procedure:

+ - **Plan mode:** Read [plan-mode.md](plan-mode.md) — locates the feature documents and selects task/target metadata for routing.

+ - **Standalone mode:** Read [standalone-mode.md](standalone-mode.md) — selects request/target metadata without investigation.

+ Scope routing returns filenames/metadata only. For symbol lookup use a filename-only search (`Grep` with `output_mode: files_with_matches`, or `rg -l`), never matching source lines. In plan mode extract only the task headings/status/dependency/target-path fields specified by the mode procedure; do not use a full `Read`/`cat` of `tasks.md` as metadata, even when it is short. If a target is unclear, route a conservative file list rather than investigating it early.

+

+ 1. **Detect and load profiles.** Run `shared-profile-detection.md` against candidate target filenames and any diff-so-far, however small the task. Read every Known profiles detection rule in a batch before choosing active profiles. Reconcile the set/count of returned rule-file paths against all expected paths, not the set of matching profiles: a nonmatching rule still must be read, never inferred from an extension. Resolve the plugin root from the loaded skill location (parent of `skills/`) or the hook-exported root; do not use profile-directory discovery as a substitute for the Known profiles list. For each active profile, read `${TOOLBOX_PLUGIN_ROOT}/profiles/<name>/implement/index.md`; ENOENT means no implement guidance for that profile. Read every always-load, matching and conservatively selected instruction. An index or requested read alone is not a loaded checklist. The next action after those returns is to emit the checkpoint below, naming returned paths and each conditional decision. No investigation may share a call with these instruction reads or precede that checkpoint. A failed/missing read keeps this gate closed; an empty profile set skips content loading, never detection.

+ ```text

+ Instruction-loading checkpoint:

+ - Shared protocols and selected mode: [returned paths]

+ - Detection rules: [expected and returned paths reconciled]

+ - Active profiles and implement instructions: [returned paths; conditional decisions]

+ - Unread required instructions: [none, or stop and resolve]

+ ```

+ 2. **Load applicable dependency instructions.** For an introduced/changed dependency — import, version, unfamiliar call, Kubernetes API, CRD, Helm chart/dependency or container image — apply `/kk:dependency-handling` before writing the call. Use its lookup cascade and the active profile's `overview.md`; do not guess signatures or configuration. Revisit this gate if investigation discovers a dependency change.

+ 3. **Investigate and establish change context.** This is the common investigation entry point for both modes. First load full requirement context: in plan mode read the entire `tasks.md`, `design.md` and `implementation.md`; in standalone mode examine the request and any referenced issue (`gh issue view`). Search the mode's declared Capy sources for relevant context, even for trivial work. Then read relevant source/current diff, trace affected entry points and unchanged callers through state and response contracts, and reproduce the problem where useful. Apply the shared change-context contract: establish intended outcomes and their authority, preserved behavior, actual scope/baselines and applicable delivery constraints before implementation edits. On resume, compare current repository state with earlier context and supersede stale evidence. If requirement loading or source exploration adds a target or reveals a new profile/conditional, pause that investigation to re-detect and load its guidance before proceeding.

+ Check whether the proposed approach satisfies the requirement and delivery policy. A concrete conflict requires the affected flow, trade-off and smallest viable alternatives before dependent edits; pending future work is no excuse for breaking a current flow. If explicit requirements conflict and no existing accepted decision settles them, record the alternatives as **proposals** and stop dependent edits until the user resolves the conflict. Do not invent document precedence, label your preferred interpretation an accepted decision, or weaken one requirement merely because the alternative preserves compatibility. Reuse actual accepted decisions and authorization; choices within those requirements create no new approval gate. State a brief approach for non-trivial work. Keep trivial changes proportionate, without irrelevant release gates or duplicate tests.

+ **Check the proposed resolution against a concrete boundary case.** For each conflicting condition, compare the outcome it requires with what the alternative would do. If the alternative permits an outcome an explicit condition forbids, it changes that requirement. In particular, replacing an unconditional guarantee/rejection with conditional or best-effort behavior is not merely an interpretation, even when it satisfies another hard requirement. Record that counterexample and wait for an actual user decision before dependent edits; saying the alternative “honors both” or assigning precedence to a document is not authorization.

+ In plan mode, set the task to `in-progress` and record observations under its labeled **Execution context**, following [plan-mode.md](plan-mode.md). Standalone context remains conversational unless material work must be deferred.

+ 4. **Implement.** Follow the established requirements and plan, if present. An observation or passing test does not authorize rewriting a requirement. Keep task/subtask progress accurate as work completes.

+ 5. **Verify outcomes.** Invoke `/kk:test` and load its resolved instructions before running verification; running a test command alone does not execute that skill's workflow. Use relevant focused checks for intended behavior, preserved contracts and applicable compatibility cases. Record commands, actual results, baselines and limits. Green tests support only the scenarios exercised; they do not prove deployment readiness.

+ ### Step 3: Report and Review

+ 1. **Refresh the handoff.** Reconcile current diff scope, candidate state, requirements, baseline identities and verification references against the repository after edits, fixes or resume. Obtain the actual selected Git diff/current state; do not synthesize diff hunks from remembered edits when a command fails. Retry an authorized read directly or report the specific unavailable evidence. Replace stale claims with dated evidence. Include task scope (done/current/pending tasks in plan mode), factual change context, affected paths, attributed test results and explicit unknowns. Do not pass implementation-session history, author conclusions as certification, or a preselected verdict.

+ 2. **Review.** Invoke `/kk:review-code` in isolated mode (or the registered `/kk:review-code:isolated` command where available) and follow its full instruction-loading workflow. If the harness has no invocation tool, read that skill's entry point and isolated procedure explicitly. Do not replace it with a direct code-reviewer spawn: the isolated workflow owns both independent-agent and PAL paths, resolved criteria and evidence preparation. Supply the refreshed context and evidence to that workflow, which prepares the same relevant source/historical evidence for both reviewers and handles further evidence requests; do not issue a duplicate PAL review. If the user selected standard review, supply the same factual record to `/kk:review-code`. Missing specification or deployment evidence narrows conclusions; it does not reduce review to code quality alone.

+ 3. **Resolve and reverify.** Apply required corrections, refresh affected context and tests, and obtain review of the changed evidence before finalizing. Index new non-obvious project conventions if any; review skills own their findings indexing.

+ 4. **Assess completion.** Report requirement coverage, verification and review outcomes, with code completion separate from release, migration or activation conditions. A known unmet hard task/delivery requirement prevents `done` unless explicit user authorization changes it or accepts an exception; record the source, rationale and prerequisites. Unknown evidence essential to acceptance remains open. An external deployment prerequisite may outlive code completion only when consistent with the agreed contract; describe release readiness as conditional or unknown.

+ Record every deferred material action durably with owner (or explicitly unassigned), reason, concrete next action and verification condition: the current task's Execution context in plan mode; existing tracking or a concise repository-local review note in standalone mode. A final-chat caveat alone is insufficient. Only then mark the plan task `done` if the contract and all **Required Outputs** are satisfied; otherwise keep required work open. Do not continue with incomplete outputs.

+

+ - An external blocker cannot be resolved within the authorized scope

+ Ask for the missing decision or capability rather than inventing a requirement. Continue independent authorized work while dependent work waits; ordinary fixable test failures should be corrected and reverified.

+ - Follow the agreed plan; record explicit authorization for any requirement change