Bugfix: adopt an unconfigured repository before the session is spawned

A bugfix in place of requirements.md (both clear the same gate, decision-045): the behaviour was specified and built in issue-193; only its ordering is wrong. The design lives here too rather than in a design.md of its own — the fix is one new method, two call sites and a test that asserts a sequence.

Introduction

issue-193 made the-loop write its built-in default harness config into a repository that carries no .the-loop/. It is programmatic and on the daemon path — and it runs on the wrong side of the only moment that matters. Ticket: #201.

sequenceDiagram
    participant D as Dispatcher._spawn_session
    participant W as Workspace
    participant G as GraphLink
    participant T as tmux + claude
    D->>W: _prepare_workspace()
    W-->>D: cwd (the checkout exists)
    Note over D,G: issue-201 writes the config HERE
    D->>G: context() — read-only
    D->>T: tmux.spawn() — the harness starts
    Note over D,G: issue-193 wrote the config HERE — too late
    D->>G: on_spawn() → _adopt()

The window between tmux.spawn and on_spawn is short and the write is fast, but a harness whose whole claim is predictability cannot leave "did the session find a config?" to a race with its own SessionStart hook. The bug is not that the outcome is wrong — the file does appear — it is that nothing guarantees when.

Requirements

Requirement 1 — the config exists before the harness starts

User story: As an operator whose daemon spawns sessions into fresh clones, I want the harness config on disk before the harness process exists, so that a session can never begin in a repository with nothing to read.

Acceptance criteria (EARS)

1. WHEN the dispatcher spawns a session for a work item in a checkout carrying no harness config THEN .the-loop/harness-config.yaml SHALL exist before tmux.spawn is called.
2. WHEN a dead session is respawned THEN adoption SHALL run in the same pre-flight as the harness-trust preparation, before the harness process restarts.
3. WHEN the ordering is verified THEN it SHALL be observed from inside the spawn call, not after the dispatch returns — the outcome was already correct before this fix, and only the ordering was not.

Requirement 2 — the gates do not move with it

User story: As the same operator, I want the earlier write to be no less careful than the later one, so that fixing an ordering bug does not widen what the daemon may touch.

Acceptance criteria (EARS)

1. WHEN adoption runs pre-spawn THEN it SHALL apply the same gates the driving actions apply: the coupling is enabled, the work item is nameable, an authorized human armed it, and the checkout is proved by its origin remote to be the work item's own (issue-113 A6, decision-044).
2. WHEN the work item walks pdlc-contribution-loop THEN nothing SHALL be written, exactly as before (issue-185, PR #187).
3. IF the checkout already carries a harness config THEN it SHALL be left byte-for-byte as it is.
4. WHILE the graph's driving actions run, adoption SHALL remain available as an idempotent safety net, so a session spawned before this change — or reaching the graph by a route the dispatcher does not own — is still adopted.

Security considerations

No new attack surface, and one gate order preserved deliberately. The write target, the allow-listed owner/repo substitution, the symlink containment (_inside) and the never-overwrite rule are all unchanged — this work item moves when scaffold() is called, not what it does.

The one thing that could have gone wrong is the reason GraphLink.adopt exists at all rather than a bare harness_config.scaffold(cwd, …) in the dispatcher: at the pre-spawn point, cwd is not yet known to be the work item's repository. Under the legacy spawnWorkdir setup it is a static directory that may be the operator's own checkout, and writing there would be issue-113's A6 defect with a different payload. So the ownership proof moves with the write.

• Abuse cases (EARS):
  1. WHEN a spawn is prepared in a checkout whose origin is not the work item's repository THEN nothing SHALL be written (test_a_foreign_checkout_is_never_adopted covers the same gate on the _guarded path; the pre-spawn path shares the method).
  2. WHEN the work item is a contribution THEN nothing SHALL be written.
• Fail closed: every gate that cannot be evaluated returns without writing.

Out of scope

The content of the default, the substitution rules, and the harness.config_scaffolded event — all settled in issue-193 / decision-073 and unchanged here.

Testing

Row	Type	What it proves	Command
T1	Integration (ordering)	.the-loop/harness-config.yaml is present at the moment tmux.spawn is called, asserted from inside a FakeTmux.spawn override; verified red by disabling the pre-spawn call	pytest cli/tests/test_harness_config_scaffold_integration.py -k before_the_harness
T2	Integration (sequence)	the dispatcher's recorded call order is adopt → context → deliver → spawn	pytest cli/tests/test_graph_drive_integration.py
T3	Regression	issue-193's whole suite — gates, carve-out, abuse cases — still passes with adoption moved	pytest cli/tests/test_harness_config_scaffold_integration.py cli/tests/test_harness_config.py
T4	Regression (whole suite)	nothing else moved	make test
T5	Lint / type-check	repository gates	make lint format-check typecheck validate

Verification results

Activity	Command	Outcome	Evidence
T1	pytest … -k before_the_harness	pass — and fails with the pre-spawn call removed, which is what makes it a test of the ordering rather than of the outcome	evidence/ordering.md
T2	pytest cli/tests/test_graph_drive_integration.py	pass — 9 passed	evidence/ordering.md
T3	pytest cli/tests/test_harness_config*.py	pass — 47 passed	evidence/ordering.md
T4	make test	pass — 1782 passed, 1 skipped	evidence/ordering.md
T5	make lint format-check typecheck validate	pass — ruff clean, 0 markdown errors, pyright 0 errors, 7 configs VALID	evidence/ordering.md

Not executed: none.

Review comments

Appended by the-loop's record-feedback hook when a human gate approves with comments (issue-109).