Roadmap and Agentic Workflow

Roadmap and Agentic Workflow

The roadmap is a live dependency graph, not a flat backlog. Every managed issue has one milestone, one status/priority/kind/risk, at least one area, one concurrency declaration, test/evidence criteria, documentation impact, and GitHub-native blocker relationships. Native dependencies are authoritative; each issue body mirrors them for humans.

Four phases

Phase	Outcome	Live milestone	Current state
1	Complete executable self-hosted behavior over the minimal kernel	Phase 1	Complete: the generalized evaluator and self-hosted all/any/none/chain/gate are covered by the reviewed executable conformance manifest through #20
2	Make coding-agent evidence reproducible and executable in-session	Phase 2	Complete: verifier runner #23, evidence mapping #24, controller #25, registered repeats #26, in-session adapters #27, and the immutable evidence audit #28 are reviewed and merged
3	Compose and enforce monotonic organization/user policy	Phase 3	Complete: restrictive composition, enforcement, conformance, and scoped expiring waiver application are reviewed and merged through #40
4	Add safe custom syntax and presentation layouts	Phase 4	Complete: #41 through #49 are reviewed, merged, and covered by the executable completion audit

Completed delivery and evidence-generalization roadmaps

Priority is considered only among issues with no open native blockers and an available concurrency token.

• The original four delivery phases are complete through the Phase 2 evidence audit #28; every issue in milestones 1–4 is closed and marked done.
• The separate Future research — Evidence generalization milestone is also complete without expanding those delivery claims. Reviewed decision #91 froze the repository populations, exact artifacts, one model, arms, 36-session schedule, analysis, exclusions, stopping, billing preflight, and zero-incremental-spend resource budget before any new model turn.
• Positive in-session use #92 completed all 12 sessions with 0/12 registered use and 0/12 acceptance. The paired BHCP-versus-prose/calibration study #93 completed all 24 sessions with 0/12 acceptance in each arm, paired risk difference 0, exact McNemar p=1, and 12/12 conservative claim calibration in each arm. With no discordant outcomes, this is a failure of both representations to induce a repair in the frozen setup—not evidence of equivalence or a general representation effect.

The complete delivery-safety foundation, independently checkable self-hosted goal algebra, six frozen coding-agent pilots, valid five-session repeat, capability-bounded verifier runner with deterministic evidence mapping, fail-closed experiment controller, executable scoped-waiver contract, audited safe-profile contract, and the bounded 36-session evidence-generalization record are merged. The five milestones are complete; broader language/runtime construction and stronger external-validity studies remain future work.

Status and dependency rules

• status:ready: open and every native blocker is closed.
• status:claimed: authoritative issue/resource refs were atomically acquired.
• status:blocked: at least one native blocker is open.
• status:review: implementation is delivered and locks remain held through review.
• status:done: merged, verified, and closed.
• status:stale: a claim needs evidence-based recovery; it is not permission to steal a lock.

Blocked issues must not be started early. Closing a blocker promotes a dependent issue to ready only after its remaining blockers and resource capacity are rechecked.

Atomic claims

Labels, assignment, and comments are metadata, not locks. Cross-machine exclusion uses unique remote Git refs:

• refs/heads/codex-locks/issues/<number> — one worker per issue.
• refs/heads/codex-locks/mutex/<resource> — exclusive shared resource.
• refs/heads/codex-locks/semaphore/<resource>/<slot> — bounded capacity, slots 1..N only.

The coordinator creates all required resource refs and the issue ref atomically, then changes the label, assigns the worker, and comments with agent/task, timestamp, branch, worktree, scope, ref paths, exact SHAs, and unique nonces. A worker keeps those refs through review.

Release is a compare-and-delete fenced by the acquired SHA. If the lease fails, ownership was lost: do not delete the newer ref. Stale recovery requires inspecting the claim comment, task, branch, worktree, PR, and recent activity, then documenting the decision.

Delivery loop

1. Select the highest-priority ready issue whose concurrency token is available.
2. Atomically claim it and make the metadata match the lock.
3. Work red-first where practical; keep Rust models, CDDL, SEMANTICS, README, conformance fixtures, and wiki claims aligned.
4. Run formatting, Clippy with warnings denied, all tests, release build, and issue-specific evidence.
5. Open a linked PR; move the issue to status:review but retain all locks.
6. Independently review the exact head SHA. Queue or merge only that green, unchanged head with --squash --match-head-commit <reviewed-head-sha>; GitHub rejects author approval and failed required checks keep auto-merge blocked.
7. Close the issue, mark status:done, promote newly unblocked dependents, and release only the exact held refs with fenced deletion.

The reusable project loop is now repository-configured through AGENTS.md and .codex/project-profile.md. Those files are the executable operating contract; this page summarizes them, and live native issue dependencies remain authoritative.