Execution Log: the operator chooses how many sessions a work item's pull requests get

Append-only log of progress for the user's visibility.

Phase transitions

Phase	Entered	Reviewed/approved by	Notes
phase-selection	2026-08-17	@MadaraUchiha-314	Declared by the owner filing the ticket and pointing this cloud session at a designated branch. brainstorming skipped — the ticket names the option it wants and the alternative was already argued in decision-088. design-critic-review not selected: no critic is configured in this repository (reviews.critics: []). See Deviations from the standard gates.
requirements-definition	2026-08-17	pending — PR for this branch	requirements.md, three requirements. The ticket has a title and no body, so the reading was posted on the ticket before any file was written (comment) and the requirements link back to it.
design	2026-08-17	pending — PR for this branch	One key grows a value; one seam grows a requirement. Six alternatives recorded as rejected in decision-092, each with why it fails.
test-planning	2026-08-17	pending — PR for this branch	12 rows, 6 in scope; every n/a carries a reason.
tasks-breakdown	2026-08-17		12 tasks, two independent red roots.
implementation	2026-08-17		TDD: the red run captured and committed before the fix.
verification	2026-08-17		Every planned activity ran. One replan, recorded below.
needs-review	2026-08-17
complete

Pull requests

PR	Scope / tasks	Status
this branch's PR	The whole work item — the spec chain, the change, the docs.	open

Progress entries

2026-08-17 — read the ticket, then decision-088

• Phase: requirements-definition
• Did: the ticket is one sentence, and the option it asks for looked like one that already exists (routing.tmux.sessionPerPr). Read issue-253 and decision-088, merged the day before, and found that D1 had turned the same-repository half of that option into a rule — deliberately, with the reason written down: "a knob for it would be a documented way to reproduce the bug."
• Decided: this ticket is its author overruling D1 and D4 of a decision they accepted yesterday. That is their call to make, so the work is to build the knob safely, not to re-argue 088. Posted that reading on the ticket before writing anything, with the option table showing that both current values collapse the case being asked about.
• Next: separate what 088 established into the claim that is negotiable (which pull requests are candidates) and the claim that is not (two conversations never share a tree).

2026-08-17 — the invariant, and the trap under it

• Phase: design
• Did: traced _endpoint_for → _spawn_endpoint → _endpoint_cwd → Workspace.prepare and found the design already had the shape needed: 088 D2 put "a session needs a tree" at a seam of its own. Only _endpoint_for's answer was hard-coded.
• Found (the important part): naïvely making the collapse optional would have shipped a new wrong-tree bug. ensure_worktree swallows a failed branch checkout and falls back to a detached default-branch worktree. For a same-repository pull request the work item's own session already holds that branch, so git worktree add -B fails every time — and the fallback returns a distinct path, which sails past _endpoint_cwd's existing _same_path guard. The-loop would have announced a session for pull request #N sitting on main.
• Decided: Workspace.prepare(require_branch=True), passed for a same-repository endpoint only. Recorded as decision-092 D4, and it is why always is honestly documented as needing strategy: clone.

2026-08-17 — red, then green

• Phase: implementation
• Did: wrote the failing tests first and committed them alone (3a71828): 28 failed, 173 passed across test_routing.py, test_workspace.py and test_configschema.py. Evidence in evidence/red.md.
• Then: the three modes and their two derived questions in TmuxConfig; one clause in _endpoint_for; require_branch through prepare / ensure_worktree / ensure_workitem_clone / _prepare_workspace / _endpoint_cwd; the schema leaf.
• Note on the schema: the design first said anyOf. configschema.py is a hand-written validator whose supported-keyword set is asserted by a test, so anyOf would have meant implementing a combinator to say what type: ["string", "boolean"] + enum already says. Changed to the union, and the design records why.
• Fixed while green: three call sites outside the diff's subject — StubWorkspace.prepare in test_trust_integration.py, a _prepare_workspace lambda double in test_routing.py, and two TmuxConfig(session_per_pr=False) constructions that pyright rejected once the field became a str. The last two now say "never", which is what they meant.

2026-08-17 — verification

• Phase: verification
• Ran: every activity in testing-plan.md. Whole suite 2,308 passed / 1 skipped; make lint (776 markdown files, 0 errors), make format-check, make typecheck (0 errors), make validate (7 configs valid), docs↔schema parity green.
• Replanned, with the reason: T2 was planned to run in test_webhook_routing_integration.py. Its server_factory builds no git origin, so the scenarios would have needed the make_origin helpers duplicated into it. They run in test_workspace.py instead, beside test_a_cross_repo_pr_endpoint_spawns_in_its_own_checkout, which is the same shape: dispatcher in-process, faked tmux, real git. They keep their Gherkin docstrings. The matrix and the environment section were updated to say so rather than left describing the plan that was not executed.
• Corrected: the plan's Verification-results row said "13 failed" for the red run before the run existed; the actual count is 28, and the file now says 28.

Capability docs

• docs/capabilities/webhook-triggers.md — the routing behaviour is restated as the invariant (a session needs a tree, in every mode) plus the three-mode choice, replacing the two bullets that stated the collapse as a rule. A history row for issue-258 was added.
• docs/capabilities/process-graph.md — which pull requests have a session to drive an inner pdlc-pr-loop from is now the mode, not "cross-repository only".

Documentation

• docs/config/cli/routing-options.md — the tmux.sessionPerPr section rewritten for three values, with the mode table, the always + strategy: clone obligation in a warning box, and an upgrade note for the booleans.
• docs/cli/state.md — what a pullRequests[] entry means now that three modes decide whether it carries a conversation.
• .the-loop/cli-config.yaml and skills/the-loop/templates/cli-config.yaml — the key is now stated explicitly with its three values commented, in this repository's own config and in the one /the-loop:init scaffolds.
• cli/the_loop/eventlog.py — the session.pr_spawned and session.pr_session_declined descriptions (read by the-loop events --explain) say which pull requests reach them under which mode, and name the new decline case.
• cli/the_loop/graph/pdlc-pr-loop.yaml — the header comment, which told the reader an inner loop is a cross-repository thing.
• README.md and the skill's reference/ docs: not changed, and here is the reason. Neither describes sessionPerPr or how many sessions a pull request gets — the CLI daemon's routing configuration is documented in docs/config/cli/, which is where the change landed. A blank would not have been an answer; this is the answer.

Decisions

• decision-092 — How many sessions a work item's pull requests get is the operator's choice — the tree is not. Refines decision-088 D1 and D4; D2, D3 and D5 stand and are what make the choice affordable. Numbered 092 because 089–091 were taken by work merged on 2026-08-16.

Deviations from the standard gates

• phase-selection was not posted as a checklist on the ticket. This work item runs in a one-shot cloud session, not under the CLI daemon, so there is no session to wait for a reply into. The owner's own act — filing the ticket and pointing a session at a designated branch — is the declaration, and the phases actually walked are listed above. The reading of an otherwise-empty ticket was still posted first, so the gate's purpose (a human states what is wanted before code is written) is served by a comment rather than by a checklist.
• Critic review not run. reviews.criticReviewCount is 3, but reviews.critics is empty in this repository — there is no second harness configured to run one. Self-review was run and its findings are in the diff (the anyOf → type+enum change, the require_branch narrowing to same-repository only, and the None-is-not-a-typo case in the fail-closed test all came out of it).
• Human sign-off is pending, as it must be. Risk tier 4 — the diff touches .the-loop/cli-config.schema.json, a sensitivePaths entry — so autonomy.tiers."4" is human-approves-pr, and security.review.humanSignOffMinTier: 4 additionally requires a named human security sign-off. Both are requested on the pull request. Nothing here completes autonomously.