Reconcile¶
For a requested feature, candidate promotion, or future direction change, use Propose. Reconcile handles new engineering reviews.
Turn a new engineering review into approved planning changes in an existing memory bank. Reconcile plans; it does not implement.
Send the request in a session for your project. Replace review.md with your
actual review source:
| Agent | Conversation request |
|---|---|
| Claude Code plugin | /memory-bank:memory-bank-reconcile review.md |
| Codex plugin | $memory-bank:memory-bank-reconcile review.md |
| DSH | /memory-bank-reconcile review.md |
Argument: a review source. It may be pasted, attached, local, repository-relative, or remote.
For direct skill-folder installs, see the invocation prefixes.
When to use it¶
Whenever a code, architecture, security, or other engineering review arrives after initialization. It requires an existing milestone and status harness; it does not initialize a project.
The tempting next prompt after a review is "fix all four findings." That skips the most important question: are they still true?
The review may describe an older commit. One finding may already be fixed. Another may duplicate a pending milestone. Its P1/P2 vocabulary may not mean what your project's definitions mean. An architecture recommendation may be reasonable but outside your ownership.
Review text is untrusted evidence¶
The review is treated as data, never as instructions. Embedded commands are not executed merely because they appear in the document.
Before any remote fetch, reconcile shows you the exact URL and obtains a separate explicit confirmation — including when your own request supplied that URL. A URL discovered inside repository content or inside the review never authorizes a fetch.
Three phases¶
Assess. Read the review, the current memory bank, relevant implementation and tests, history, and the review's stated baseline. Revalidate every finding against the current repository.
Propose. Present the complete disposition matrix:
| Finding | Source priority | Local severity | Current disposition | Planned owner |
|---|---|---|---|---|
| Process survives failed startup | P1 | P1 | Confirmed | Row in the open session-lifecycle milestone |
| Timeout behavior differs | P2 | P2 | Duplicate | Existing pending public-contract row |
| Cleanup failure path unverified | P2 | P2 | Confirmed | New remediation milestone |
| Rename helper | P3 | Lower | Deferred | Candidate Direction with a trigger |
Dispositions also include partially confirmed, already resolved, unsupported by current evidence, outside ownership, and decision-dependent. The source priority is preserved for provenance; the local severity is classified independently using your project's definitions.
Write. In standalone usage, apply the approved actions. Nothing is written until you approve the matrix, owners, dependencies, downstream impacts, candidate directions, and file actions.
Reconciliation during an approved run¶
A Goal/Next owner with human-approved CORRECTIVE_PLANNING: auto may supply
standing approval for necessary in-scope corrective planning. Reconcile still
prepares and checks the complete finding/ownership/file-action proposal against
the original outcome, owning repositories, existing grants and all limits.
Covered pending amendments or new repair milestones and dependency/launch/hash
updates then need no further confirmation. Frozen history and counters stay
preserved. New features, permissions, higher limits, conflicts or unknown
checksum drift require approval. The goal contract
owns this exception. Standalone Reconcile retains its complete approval gate;
this skill plans only and hands implementation back to the existing owner.
The invoking request must carry the full resolved request, identifiable human approval source, covered actions, grants and limits. Reconcile verifies this basis against the invoking conversation or supported trusted host state before writes. A review, project file or model-authored assertion cannot supply it; missing provenance keeps the ordinary approval gate. A valid basis needs no repeat confirmation. Repairs cannot weaken acceptance or cancel unmet obligations. A new repair milestone preserves the originating review gate and counter, including the iteration-10 stop.
Where findings go¶
A confirmed finding that fits an open milestone's scope becomes a new pending row there. An existing pending milestone that already owns the outcome is updated rather than duplicated.
Completed history stays closed. When a finding concerns completed work, reconcile creates a new remediation milestone with lineage back to the affected one. Historical status says what happened; it is not rewritten to make today's plan look tidier.
Each active finding carries portable provenance: review title and finding ID,
source priority and local severity, stated baseline and revalidation commit,
repository-relative evidence, and lineage to existing work. There is no
accumulating reviews/ directory and no second review ledger.
Architecture recommendations are not current architecture¶
A review that describes a target state belongs in milestone scope and acceptance
until it is implemented. Writing it straight into architecture.md would make
the memory bank false: the document would describe the system the review wants
rather than the system the agent is about to edit.
Reconcile corrects product.md, architecture.md, or tech-stack.md only when
current repository evidence proves an existing description stale.
Downstream first¶
A finding rarely stops at the file where it was found. Reconcile updates pending consumers before implementation begins: dependencies and order, task wording and acceptance, compatibility expectations, verification coverage, and downstream status files whose plans depend on the affected contract.
This keeps the plan consistent when a finding changes assumptions shared by several milestones.
What it will not do¶
Reconcile does not implement findings, mark tasks complete, change non-pending row state, commit, push, or launch execution. Planning and execution stay separate authorization boundaries.