Reject a malformed dispatch brief before it can launch.

Brief-linting

Intent — Statically lint an agent's task brief before the agent is spawned, refusing to launch any brief missing the markers that make the agent's work safe and well-scoped.

Motivation — the failure it kills

A brief is the agent's entire world: it is the only instruction set the agent reads before it starts mutating a repository. A brief with a missing marker does not fail loudly. It fails silently and downstream. Omit the worktree-isolation marker and the agent edits main directly, with no fence firing. Omit the dispatch-id and the lifecycle substrate can't track or clean the agent. Omit a mandatory safety snippet (PATH export, commit-cadence, submodule check) and the agent trips a sharp edge 20 minutes in. Because brief authoring is a manual act repeated for every dispatch, the failure is not a one-off; it is a class that recurs on each launch and compounds across a concurrent fleet.

Why it's not just "review the brief before you send it"

Human (or orchestrator) review of a brief is a probabilistic check: it catches the missing marker if the reviewer happens to look for it, and reviewers reliably miss the boring structural markers precisely because they are boring. Brief-linting is a deterministic check at the point of no return. It greps for each required marker string and exits non-zero if any is absent, so a malformed brief cannot be dispatched rather than is unlikely to be. It also moves the check to the one moment that matters: the pre-dispatch instant, after which the agent is autonomous and the cost of the omission is unrecoverable. A gate catches the boring marker every time; review catches it only when the reviewer thinks to look.

Mechanism

The brief linter runs a battery of grep-style checks over the brief text: worktree-isolation marker present, dispatch-id: agent-<id> token present and the matching per-agent marker file exists on disk, every mandatory snippet's marker string present, subagent_type declared. Exit 1 ⇒ do not launch. It is not a standalone habit: the canonical dispatch path calls it as a required prepare step and rolls back the just-written marker file if the lint fails, so the dispatch path is the enforcement.

Adopt it — the agent-brief starter is the portable shape a brief takes (Epic cite · trigger · role/LLM · time cap · scope allowlist + non-goals · pre-pasted file:line context · verification · acceptance · trust-nothing hand-back). It is the third of the three artifacts an orchestrator authors: the Epic is the effort, the design doc is the plan, the brief is the dispatch. The marker/snippet plumbing this lint checks is the harness-specific enforcement layered on top.

Genre-gate the content checks (so the lint can grow without crying wolf)

Marker-presence is the structural layer. A mature brief lint adds a content layer (does the brief cite a real file:line, declare its footprint, cite a reachable commit SHA?), and that layer has a trap: a content check that fires on any mention (a filename, a path, a SHA-shaped token) false-positives on briefs where the citation isn't required. A doc-only brief that merely names a source file trips the "cite the code you touch" check; the operator learns the lint cries wolf and starts ignoring its output, the exact failure the whole mechanism exists to prevent.

The fix: the brief declares its genre. It carries a brief-genre marker (emitted by the same template that lays down the structural markers), and each content check gates on it, firing only when the genre requires that citation (the "cite the code" check on a code brief, not a design-doc one; the SHA-reachability check only where a citation is actually required). No genre marker → every check fires, the safe default. Genre-gating is what lets the content layer keep growing (more checks, more precision) without eroding the trust the structural layer earned. A lint tuned out is worse than a check not written.

Prerequisites

• Briefs are structured text with grep-able marker strings, not free prose; the lint's checks are string presence assertions.
• A mandatory-snippet registry the lint can enumerate (see mandatory-snippet-table).
• A dispatch wrapper that calls the lint on the canonical path and refuses to proceed on failure. Otherwise the lint is optional and therefore skipped under time pressure.

Consequences & costs

• Structure is not correctness. The lint checks that a brief is well-formed, not that it is well-scoped: a brief with every marker present can still task the agent with the wrong thing. A green lint buys safety-marker coverage, not a good brief; the human still owns scoping judgment.
• Authoring tax. Every brief must now carry boilerplate markers, making hand-authoring heavier. A brief-template generator mitigates this by pre-placing them by construction. Without the generator the markers become friction the author routes around.
• Each new marker is a maintenance edge. A new mandatory snippet means both a new lint check and threading the marker into the template; drift between the two produces false rejections.
• Bypassable by design. ADA_TOOL_BYPASS_AGENT_FENCE=1 exists for humans; the mechanism's floor is the discipline of not misusing the escape hatch (audit-logged, but still a hole).

Known uses

• The agent CWD-drift-defense rule: the isolation: "worktree" marker is BLOCKING and lint-verified.
• The dispatch wrapper's prepare step emits its go-ahead only after the brief passes.
• A dedicated check verifies the dispatch-id token and the on-disk marker together, closing the "token present but no live agent" gap.

Related mechanisms

• Enabler — mandatory-snippet-table supplies the snippet markers this lint asserts; the registry must exist before the lint can check for them.
• See also (complement) — dynamic-context-injection: the content side of dispatch (injects the rules relevant to the target files) where brief-linting is the structure side (checks the brief is well-formed). Same stage, orthogonal jobs.
• See also (complement) — role-typed-dispatch: the other half of a well-formed dispatch, the role that fixes LLM, isolation, and gate set.