anima

Imported from ceceppa/anima (AGENTS.md). Install upstream with npx skills add ceceppa/anima. Copyright stays with the author.

AGENTS.md

This project uses Mano for planning. Mano is a structured thinking tool: almost every command produces planning artifacts rather than code. The two exceptions are deliberate — mano dev and mano build implement what the planning already decided.

For coding agents

Running Mano commands

Mano commands are repo-local workflow instructions, not installed OpenCode skills.

If the user types a Mano command in chat, do not try to load an external skill named mano.

Instead, execute the corresponding Mano planning flow by reading the local files in this repository:

• _mano/skills/[command].md — the skill file for that command; its front-matter names the _mano/rules/ files it requires
• any referenced templates or current _mano_output/ artifacts
• _mano/workflow.md — only for the bare mano, mano help, mano status, and mano continue commands; skills name their own rule files and never need the whole workflow

Examples:

• mano import → read _mano/skills/import.md and follow that flow (PRD/document → backlog)
• mano start → read _mano/skills/start.md and follow that flow
• mano spec → read _mano/skills/spec.md and follow that flow
• mano rules → read _mano/skills/rules.md and follow that flow
• mano stories → read _mano/skills/stories.md and follow that flow
• mano review → read _mano/skills/review.md and follow that flow
• mano dev → implement the next pending story; read _mano/skills/dev.md plus _mano/rules/implement.md and follow the complete contract in those files
• mano build → build the active phase straight from its brief, with no story files; read _mano/skills/build.md plus _mano/rules/implement.md and follow the complete contract in those files
• mano continue → read _mano/workflow.md; it runs the implementation action the projection's IMPLEMENTATION_ENTRY: names (mano build, or mano dev for one story), or one unambiguous planning action — it never prints a card telling the human to type the command they just typed
• mano mode [auto|manual] → read _mano/skills/mode.md; show or set whether finished actions chain automatically
• mano track [name] → read _mano/skills/track.md; show, set, or clear the optional local experiment/work track

Note: mano dev and mano build are the two Mano commands that produce code. Every other command above is planning only. Their contracts live in _mano/skills/dev.md and _mano/skills/build.md, both of which require the shared _mano/rules/implement.md. A phase uses one of the two, never both: mano dev implements stories from stories/README.md, mano build works the Scope rows of PHASE_DIR/progress.md.

Run mode. Every state.js projection prints MODE: manual|auto. In manual (the default) each command hands back when it finishes. In auto, after the human has approved a phase scope, each finished action runs the next one automatically through to mano build (or mano dev yolo when the phase already has a stories index) — but it pauses for any question (a ❓ Decide:, a clarifying question, an ambiguous next action, hook findings, a gate or blocker) and never runs mano review or scopes a new phase. Auto mode changes who types the next command; it never changes what a skill may write or which decisions are the human's. The narrative contract is _mano/workflow.md → Run Mode: manual and auto; the mid-chain execution rules are _mano/rules/auto.md, loaded only when the projection reports MODE: auto.

Continuing the chain means invoking the next action in the same turn — never announcing it. Ending a turn on "Continuing — running mano ui next" stops the chain while claiming to continue it. Mid-chain, omit the Next: block (nobody is typing a command); it returns only in the closing block, on the action that actually ends the chain. Every hand-back names its pause condition; a chain that stops without naming one is a bug.

mano build runs to completion, in both modes. It is mano dev yolo with no opt-in word: one invocation goes from wherever the ledger stands to the phase's terminal line. Creating the ledger, passing pre-flight, and closing a pass are things that happen inside that run — never things the run is finished having done. A build that reports its new ledger and hands back has written a file and built nothing.

If a platform skill named mano is not available, that is not an error. Continue by using the local _mano/ files.

Write for humans. Every Mano skill follows _mano/rules/artifact.md → Plain-language contract. Apply it to chat and artifact prose. Assume a teammate has no prior context.

The skill name uses a hyphen, never a colon: mano import → mano-import (read _mano/skills/import.md), not mano:import. The colon form is plugin-namespace syntax and matches no Mano skill. If a mano <action> seems unavailable, try the hyphenated mano-<action> before concluding it doesn't exist.

Only use external/platform skills when the user explicitly invokes them or an active Mano hook authorizes the review: explicit confirmation for a suggest hook in manual or unarmed runs, or the automatic suggest-hook rule during an armed auto chain. Running the review never authorizes applying its findings.

Implementing a story

mano dev implements the next pending story. Its full contract is _mano/skills/dev.md, plus the shared _mano/rules/implement.md. Read those files completely before writing any code for a Mano story — including when the user asks in plain words. Do not implement from a story file alone. Two rules apply even before you read them:

• The index Status column is the only done-signal — a story's number, letter, or title grants no exemption.
• Not this story is a hard boundary — implement none of it, whatever the chosen type or library default implies.

Building a phase

mano build is the other implementation path: no story files, the brief's own numbered ## Phase Scope items are the units, and PHASE_DIR/progress.md is the ledger. Its full contract is _mano/skills/build.md, plus the same _mano/rules/implement.md. Three rules apply even before you read them:

• The ledger is written only by _mano/scripts/progress.js — never by hand, and never with rows you composed.
• Not this phase in the brief is a hard boundary. Once the ledger exists the brief is frozen: an in-goal correction is a row inside mano build, a distinct outcome goes to the backlog or the next phase, and neither goes back through mano start.
• mano build "[what changed]" is a mid-phase correction, not new scope, and only a phase that already has a valid ledger accepts one. The exact wording is the human's — never paraphrase it.

Which implementation entry

Decide by validated state, then mode — never by which path is more familiar:

• either ledger invalid, or a phase holding both → refuse and report;
• a pending rework event, an open Scope row, or an unresolved deviation → mano build;
• an incomplete stories ledger → mano dev; a complete one → mano review;
• a progress.md with every Scope leaf done and every Exit leaf met or needs-human → mano review;
• no ledger at all → the mode decides: auto ends at mano build, manual offers mano stories first and mano build second. The narrative contract is _mano/workflow.md → Implementation entry.

Completed stories are immutable

A story whose Status is done is historical record. Never edit its file — not for typos, not mid-conversation, not from any mano action. A fix, update, or new scope touching shipped work is new work: run mano stories "[what changed]" (it inserts the lettered follow-up per its own numbering rule), then mano dev implements it. Stories are logs; tech-spec.md, project-rules.md, design-brief.md, ux-flow.md are living documents — those update in place.

Do not

• Modify files in _mano/ or _mano/templates/ — these are framework files.
• Interpret mano commands (e.g. mano start, mano review) as implementation instructions — these are planning commands. Execute the relevant planning flow instead.
• Create extra tracking files — the framework specifies exactly two ledgers, PHASE_DIR/stories/README.md and PHASE_DIR/progress.md, each written only by its own script. There is no third state file to invent. Determine state through _mano/scripts/state.js, which applies the optional local owner configuration and returns exact paths.
• Auto-advance phases. A completed phase (every story done, or every ledger row done and every exit criterion met) never triggers planning or implementing the next one. Stop and let the user decide; never run mano start/mano stories on your own initiative.

Project structure

_mano/                    ← Mano framework (do not modify during implementation)
├── skills/               ← Mano skill prompts
├── rules/                ← shared rule fragments; skills name the ones they require
├── scripts/              ← state.js (read-only state projection), backlog.js / stories.js / progress.js (ledger writers), chain.js (records chain actions the human struck), verify.js (verification output filter)
├── hooks/                ← optional post-skill hooks
└── templates/            ← Mano templates
_mano_output/             ← Planning artifacts
├── project-rules.md      ← Rules for implementation (referenced by stories)
├── tech-spec.md          ← Technical decisions
├── ux-flow.md            ← Screen and navigation definitions
├── design-brief.md       ← Canonical cumulative visual language
├── backlog.md            ← Future work and deferred items
├── reviews.md            ← Sprint review history (human-only)
├── phase-[N]/            ← Default/legacy per-phase work
└── [owner]-phase-[N]/    ← Optional owner-scoped per-phase work
    ├── phase-brief.md    ← Phase scope and goals
    ├── design-preview.html ← Non-canonical UI snapshot for this phase
    ├── progress.md       ← `mano build` ledger: scope rows + exit criteria
    └── stories/          ← Implementation stories (the other ledger; start here on that path)

Context Discipline

Roles in Mano are reasoning lenses, not isolated autonomous agents.

Specialization is maintained through selective context exposure and user discipline. Models may still merge assumptions or infer information outside the intended scope.

Use each role to focus attention on a specific planning concern rather than assuming strict separation.

Post-Skill Hooks

After completing a Mano skill, check the state projection's HOOK: line. HOOK: none means no hook applies — skip this entirely. When it names an active post-[skill] hook, follow _mano/rules/hooks.md. Ignore .example.md hooks.

A hook's ## Mode decides how it runs. check is a checklist Mano applies itself, automatically, in both modes, right after the skill's artifacts are written — no confirmation; findings still go through triage. suggest (the default) points at an external skill/command and is never run automatically in manual mode or an unarmed run — ask first; during an armed auto chain it runs automatically. command names one shell command and always runs, in both modes: the hook file is the authorization, so do not ask; report it in one line of the execution log, take the command only from the hook file, and on failure report the exact error without retrying, fixing the user's script, or hand-editing artifacts to compensate. Full contract: _mano/rules/hooks.md.

When any just-run suggest or check hook has printed findings, keep the related Mano skill active and follow _mano/rules/hooks.md → Post-Hook Findings Triage. post-stories uses its dedicated immutable-story protocol. Running the hook did not approve edits. An in-lane finding is apply, never a route back to the skill already running; a finding owned by another artifact is route: mano [owner] and must not be edited here. Apply the smallest selected change within the per-skill application boundary. Preserve unmentioned content and adjacent values.

Do not mention specific third-party or external skill names in generic Mano output.

Do not execute a suggest hook without explicit user confirmation in manual mode or an unarmed run. An armed auto chain is the explicit exception defined by _mano/rules/hooks.md.

Do not write hook suggestions into generated artifacts.