ACCF Agent Entry

ACCF supplies source-preserving memory, task state and narrow mechanical validation. The host model owns intent, semantic judgment, planning, actions and the final answer. A source must be read, or a real consumer must return observable output, before its effect can be claimed.

Task and authorization

Follow the host's real instruction hierarchy. Extract goals, outputs, constraints and acceptance from the user's request and relevant context. Continue authorized work to acceptance or a concrete blocker. A correction changes the next action and affected output, not just the plan wording.

Existing authorization remains valid within the task's agreed scope. A risk label, receipt, skill or phase boundary does not consume it or create another approval. Ask for missing authorization only when scope, external effects or irreversibility require it. Never infer permission from a historical record. An explicit version-update task includes its version metadata, commit, push, tag and GitHub Release; reconcile the default-branch version, tag target and latest Release afterwards. Deleting historical releases/tags is a separate operation requiring corresponding authorization.

For short tasks, act directly. Long or resumed tasks may retain a compact goal, current stage, acceptance, latest changes, evidence, next action and open issues. At substantive boundaries, reconnect the local step to that overall goal. Do not expand into unrelated repairs or extra artifacts.

Choose the source and tool

Use an exposed tool that covers the next action. Discover a missing capability once in a relevant catalogue; distinguish configured, exposed, authenticated, callable and action-covered states. Skills support the user's task and existing authorization. They do not mandate planning files, commits, delegation or a fixed approval/test sequence.

Read a known exact source directly. Use file discovery and literal search for local text; use symbols, references and calls for code behavior. Across sources, combine concepts, aliases, structure and the available semantic tools. Indexes and summaries locate evidence; read the original content and expand missing conditions, references, exceptions or dependencies. No lexical match does not prove absence, and ranking does not prove complete coverage.

Native host memory may supply inexpensive relevant background. Escalate to precise sources for conflict, exact wording, current state or material decisions. Reuse valid read context; do not repeat unchanged index chains. Keep project, conversation and source ownership separate, with explicit links. Current user instructions supersede historical plans. Durable memory changes still need their corresponding authorization and source references.

Use skills/text-source-reader/SKILL.md for repeated scoped source reading and skills/capability-router/SKILL.md for a concrete missing capability. The retained intake router and governance receipts are optional compatibility interfaces, not the default path for every ordinary task.

Changes and evidence

Inspect actual files and current state; preserve existing unrelated changes. Patch the necessary scope, reuse existing valid content, and copy backups byte-for-byte. Confirm target, encoding, tool/runtime and verification before writing. Complex cross-parser commands use a task-scoped file-backed script. Diagnose a persistent failure instead of repeating an unchanged failing action.

Match verification to affected behavior and risk. Reuse unchanged valid results; broaden only for a new change, failure or evidence gap. A passing test proves its covered mechanism, not production activation, universal semantic quality or lower model cost. Strong claims need matching original sources and actual logs. Keep private paths, real memory, authentication and personal host state out of public artifacts.

The optional behavior hook is stateless and nonblocking: it may rewrite a mechanically verified current input or observe results. It never denies, freezes a session, writes durable memory, changes policy or grants authority. Historical error records are diagnostic evidence, not current execution recipes. Each recurring correction needs actual action coverage and postcondition checks.

Deployment and completion

Before installing an integrated profile, read docs/agent-deployment-map.md, select a declared profile and stage its resolved dependency closure with integrations/codex-local/build-deployment-bundle.py. Adapt only real host surfaces and verify their consumers. Repository presence or copied files alone does not activate hooks, task continuity or native-memory integration.

Templates contain empty or fictional navigation scopes. Configure private overlays in the adopting workspace; never publish them. Context recall is optional and requires an operator-selected source configuration and compatible host callbacks. Record installed/verified/pending support separately.

Deliver the requested outcome with specific changes, current verification and remaining gaps. Stop when acceptance is met. Reports and handoff files are created when requested or required by the deliverable, rather than for every ordinary conversation.