aiimagined/godmode

Coordinate local context continuity, evidence, and guarded coding workflows. Use when starting or resuming substantive repository work, when current reality is uncertain, or when a request spans multiple Godmode capabilities. Not for a single-file edit with no continuity, evidence or gating concern.

Account for what a session or a release spent and hold it to declared ceilings: check reported tokens, tool calls and seconds against the run ceilings, read the per-session spend series with its measurement gaps, fold burn beside gate activity per session and per release, and measure the brief budgets and verb timings locally. Use when a session is burning more than it should, a release cost more password rounds or suite reruns than planned, or a ceiling needs to be checked before more work is spent. Not for detecting a repeating retry loop, and not for release notes.

Draft and verify release-notes from local commits and changelog fragments. Use when checking that every user-visible update carries a fragment, merging fragments for a version, or building and checking the release-notes for that version. Not for a one-line note authored by hand and not for a code-review summary of a single file.

Run CI's checks locally, in CI's order, before opening a pull request: the task precheck, the fragment gate against the base branch, the preflight of HEAD or a working-tree snapshot in a disposable worktree with the designated suite in shards, the install self-test and the control grid, the version reconciled across every surface, and a fresh hook proof on the host. Use when a PR is about to be opened, a CI job went red and the failing leg must be reproduced locally by its shard, or the local gate run is slower than the gate it feeds. Not for merging fragments into release notes, and not for approving a destructive action.

Use at the start of every session and before every new task - read GODMODE-CODE-OF-LAW.md at the project root when it exists (the project's generated Code of Law; godmode law compile writes it), and follow its guards before acting. Not for writing new laws; do not use when GODMODE-CODE-OF-LAW.md is absent.

Query a local, on-demand code dependency graph to answer who calls, who imports, and what blast radius a change carries, before spending an unnecessary full read on a large module, refreshing the saved graph from the archive first when it looks stale. Use before an edit whose impact or required retests are not yet named. Not for a routine read with no graph question involved, or for chasing an already-failing check.

Reconstruct and preserve private project continuity from local evidence. Use when resuming work, changing branches or worktrees, detecting drift, recording handoffs, or explaining why context is missing or stale. Not for code investigation or release steps.

Shape a claim so its citations grade the first time: reflect it against the record, cite exit-bearing commands through --verify, give a hypothesis its --refuted-by falsifier, run falsifiers that aged past due, repair stale or loose citations, and resolve the claim once with fresh evidence. Use when a claim keeps being downgraded, a Stop gate keeps blocking on unsupported sentences, or cited evidence has drifted. Not for a second audit of a pass already accepted, and not for root-causing a failure.

Preview and govern protected engineering actions without executing them implicitly. Use before destructive, externally visible, history-changing, database-changing, release, credential, branch, or worktree operations. Not for reads, tests or edits inside the working tree.

Detect every installed AI CLI host, reconcile Godmode's interception manifests into each one's native format, and drift-repair the installed configuration with a shown diff before anything trusts interception. Use when installed manifests need reconciling or repairing after an install, an update, or reported drift. Not for explaining what an event or a mechanism does, and not for authorizing an unrelated protected operation.

Score a staged commit's fan-out and architectural-boundary crossings just before committing: dependents left unfollowed, hook-to-runtime import direction, paths outside the plan's editable fence, paired artifacts left behind, and the green retests the commit-time closure refusal will demand, with a recorded decision as the named escape hatch. Use before committing a multi-module edit. Not for answering who-calls questions ahead of an edit, and not for authorizing a push or release.

Diagnose technical failures with reproducible evidence and bounded hypotheses. Use when a bug, regression, test failure, build failure, performance problem, or repeated unsuccessful fix needs root-cause investigation or a postmortem. Not for feature work or release mechanics.

Bound an iterative fix-and-retry cycle before it starts and stop it when it stops making progress: audit the loop contract, declare the iteration cap and stop conditions, tick every iteration with its evidence, read the host transcript for repeating error signatures and overlapping hunks, check whether blaming the model is supported, and close the loop as finished or cut off. Use when the same failing test is being retried, edits keep landing on the same lines, or an agent has looped past its budget without noticing. Not for root-causing the failure itself, and not for open asks that repeat across sessions.

Garden the local archive at session close or idle time: tell the session as dated prose, list near-duplicate or contradicting lessons and decisions, preview which old episodes expire into the cold tier, supersede a duplicate, and cluster correction candidates for a promotion that cites its rationale. Use when lessons pile up duplicated, stale records crowd the brief, or a candidate cluster is ripe to promote. Not for resuming work at session start, and not for scrubbing a secret.

Re-pitch an answer that did not land, and record why it did not. Use when the operator says they do not follow, asks for clarity, repeats a question already answered, or asks what is being waited on. Not for first answers; do not use when the operator has not signalled confusion.

Rebuild a mechanism already studied in a reference implementation, in stdlib under Godmode's own names, then pin it with a reproducing test proven able to fail, cite the reference receipt and the replicating test on the host row, and scrub private terms before committing. Use when a host or platform row is marked unverifiable while its reference sits unread or unreplicated. Not for a first survey of an unread codebase, and not for an interactive session proof.

Study an outside source repository in a fixed order - licence, tree, the two or three implementing source files, line-cited receipts - and record both an import verdict and a behaviour verdict, refusing adopt or extend on a surface-only reading. Use when an outside codebase is being surveyed for a mechanism worth borrowing. Not for rebuilding that mechanism under Godmode names, and not for root-causing a failure in this repository.

Independently recheck a change, session, PR, or diff that already carries an accepted claim or verdict. Use before trusting an earlier pass as settled, sorting any new finding into a gap, bug, verification-miss, or scope-drift. Not for a first-time look at work nobody has checked yet, and not for re-running the same suite the first pass already ran.

Score a drafted or hand-edited skill's authored routing rows against the committed baseline, and confirm its bundle validates, before it ships or is enabled by default. Use when a new or edited skill needs its overlap scores measured and its ratchet checked before landing. Not for authoring a brand-new skill, and not for judging whether a change is correct.

Create an original, compact project skill after a repeated capability gap is proven. Use when at least three real tasks need the same missing workflow and existing skills do not cover it reliably. Not for one-off tasks; do not use before three real tasks proved the gap.

Move a specification and its plan through specify, approve, implement, review, and status as one linked, resumable record instead of separate hand-authored drafts that drift apart. Use when a task needs its spec written, a plan approved before mutation, work bound back to that plan, and its status checked or reverted. Not for authoring a standalone document with no tracked contract, and never as a way to approve a plan against yourself.

Work the open-asks queue down to zero-or-mapped: list open operator asks and aging obligations, promote rule-shaped asks to invariants, map work asks to plan tasks, close answered or control-noise asks by their exact ask id, park the rest with a reason, and run falsifiers that aged unrun. Use when open asks or obligations pile up, phantom asks nag at every stop, or rules sit misfiled as asks. Not for pruning duplicate lessons, and not for re-explaining an answer that did not land.