Bugfix spec: remove the process runner — a stale session record silently downgrades tmux delivery to a dead process resume

Phase 1 of 3 for a bug (bugfix → design → tasks). Human approval for this tier-3 change happens at the PR (autonomy.tiers."3": human-approves-pr).

Summary

As reported, routing.runner: tmux only controls the runner chosen at first spawn. Once a session record exists under the registry, every later dispatch reads the runner from the record itself ("the session's recorded runner wins", decision-021). A record that ends up with runner: "process" — the dataclass default when the field is omitted — silently redirects every subsequent event for that work item to a headless adapter.resume(...) subprocess, using whatever cwd/harnessSessionId the record happens to carry, instead of the tmux session the operator is watching. No reconciliation compares the record against the live config, and nothing logs the disagreement: the poller reports a successful dispatch while the operator's tmux pane stays silent.

The owner's decision on the ticket (comment) resolves the bug at its root instead of patching the reconciliation gap: remove the in-process (headless subprocess) runner entirely. tmux is the only runner. With a single runner there is no per-record runner choice left to go stale, no config/record disagreement to reconcile, and no silent path for a delivery to take: every dispatch either lands in the work item's loop-<slug> tmux session or respawns one, loudly, through the existing issue-80/89/146 machinery.

Steps to reproduce

Run the daemon with routing.runner: tmux in cli-config.yaml.
Get a work item's registry JSON into a state where runner is "process" (manual edit, a copy/pasted session file, an out-of-band the-loop sessions register — which never sets runner, so the dataclass default applies — or a race between two near-simultaneous events).
Trigger a new comment/event for that work item.
The poller logs "routing X -> session Y" and records a successful dispatch, but nothing appears in the tmux session — delivery went to a one-shot subprocess resume, with no warning anywhere.

Expected vs actual

Expected: with the operator running a tmux fleet, every event for a work item is delivered into (or respawns) its attachable loop-<slug> tmux session; any inability to do so is loud.
Actual: events are silently consumed by headless claude -p --resume … subprocesses driven by a stale record; if the record's cwd points at another work item's checkout, the resumed process runs against the wrong working directory entirely, with no error surfaced.

Root cause (confirmed by reading the code)

Dispatcher._dispatch_one branches on session.runner == "tmux"; anything else falls to adapter.resume(...). Session.runner defaults to "process" (sessions/registry.py), so any record written without the field — including every record the-loop sessions register creates — takes the headless branch forever. RoutingConfig.runner is consulted only in _spawn_for, i.e. at first spawn. The defect is structural: two runners, a per-record selector with a quiet default, and no invariant tying the selector to the config. The fix the owner chose removes the second runner, and with it the selector.

Requirements

The wording of the owner's decision: "remove the whole in-process based claude/cursor run. ONLY run using tmux. remove all code and documentation related to process based runner."

Requirement 1 — tmux is the only runner

User story: As an operator, I want every daemon-driven session hosted in an attachable tmux session, so that nothing the-loop does is invisible to me.

Acceptance criteria (EARS)

AC1.1 WHEN the dispatcher spawns a session for a work item THEN the system SHALL host it in a named tmux session (loop-<slug>), with no configuration able to select a headless subprocess instead.
AC1.2 WHEN the dispatcher delivers an event to an existing session THEN the system SHALL deliver into the recorded tmux session, or take the existing respawn path when it is gone — never a headless adapter.resume subprocess.
AC1.3 WHEN the daemon (gh-webhook start / poll start) starts THEN the system SHALL always require the tmux binary in its dependency check.

Requirement 2 — the process runner's code and vocabulary are removed

User story: As a maintainer, I want the process-runner code paths gone, so that no stale record or config can route a delivery somewhere silent.

Acceptance criteria (EARS)

AC2.1 WHEN reading the harness adapter contract THEN the system SHALL expose no headless per-dispatch spawn/resume subprocess path; the one-shot invocation surface used by critic reviews (oneshot_argv) SHALL remain.
AC2.2 WHEN a session record is written THEN the system SHALL no longer write a runner field, and WHEN an old record carrying one (any value) is read THEN the system SHALL ignore it rather than branch on it.
AC2.3 WHEN routing.runner appears in an operator's config THEN the system SHALL ignore it and log a warning naming the removal (tmux-only), and the config schema SHALL no longer declare the key.
AC2.4 WHEN a session record has no live tmux session behind it — including a legacy record written by the process runner or by sessions register, which has no tmuxTarget — THEN the next dispatched event SHALL take the respawn path: resume the recorded conversation in a fresh loop-<slug> tmux session when possible, else start a fresh one, per the existing issue-89/146 rules.

Requirement 3 — documentation matches the single-runner reality

User story: As an operator reading the docs, I want no instruction that configures or describes the process runner, so that I cannot configure a mode that no longer exists.

Acceptance criteria (EARS)

AC3.1 WHEN reading the CLI config reference, the config templates, the skill reference docs, or the capability docs THEN the system SHALL describe tmux as the only way daemon sessions run; routing.runner and "process runner" SHALL appear only in historical records (decisions, changelog, past specs), which are not rewritten.
AC3.2 WHEN the affected capability docs are read THEN their behaviour tables SHALL reflect tmux-only dispatch, with a history row tracing to this spec.

Requirement 4 — the removal is proved, and stays proved

User story: As a maintainer, I want the test suite to pin the tmux-only behaviour, so that a future change cannot quietly reintroduce a silent path.

Acceptance criteria (EARS)

AC4.1 WHEN the test suite runs THEN tests SHALL cover: delivery to an existing tmux session, respawn of a legacy record without a tmuxTarget, the routing.runner ignore-with-warning, and the always-on tmux dependency check.
AC4.2 WHEN the full gate runs (pytest, ruff, pyright, markdownlint, config validation) THEN it SHALL pass.

Security considerations

Untrusted actors / trust boundary: the trust boundary between GitHub payloads and harness invocations is unchanged: nothing payload-derived reaches an argv, a path, or a tmux target. Removing the headless resume shrinks attack surface — the path where a corrupted registry file's cwd/harnessSessionId was handed to a subprocess (adapter.resume) is deleted; the remaining tmux paths keep their existing guards (_SESSION_ID_RE on resumed ids, _LOOP_TARGET_RE on signalled targets).
Abuse case: a hand-edited session record pointing cwd at another work item's checkout could previously make a silent headless resume run there. Post-change the same record routes through the tmux respawn path, which spawns in that cwd but visibly — named loop-<slug> session, announce comment on first spawn, session.respawned events. The registry stays local operator-writable state; its validation rules are unchanged.
Fail-closed: when tmux is unavailable the daemon refuses to start (dependency check) and a dispatch fails loudly; nothing degrades to a headless run.
No new attack surface: no new inputs, no new privileged operations; this change only deletes an execution path.

Out of scope

The reconciliation/sessions doctor alternatives suggested in the issue body — superseded by the owner's decision to remove the process runner.
Interactive support for the cursor harness in tmux (cursor-agent still raises UnsupportedRunnerError for interactive hosting, unchanged; it remains usable as a critic via oneshot_argv).
Rewriting historical records: decisions (e.g. decision-021), CHANGELOG entries, and past specs keep their text.
A registry migration tool: legacy records heal lazily on their next event via the respawn path (AC2.4); no eager rewrite of files on disk.

Open questions

None. The owner's comment on the ticket is the decision of record.