Decision 126: the execution log is retired; each gate keeps one record of its own under evidence/

• Status: proposed
• Date: 2026-09-14
• Deciders: the-loop (architect, engineer), reported by @MadaraUchiha-314
• Work item: issue-365

Context

docs/specs/<id>/execution-log.md was two different things wearing one name.

A narrative. A 130-line bundled template materialized into every work item; a phase transitions table; append-only progress entries; a log-entry hook appending a checkpoint at 47 node boundaries across five shipped graphs; a phase: front-matter field mirroring the ticket label; and a prose entry — Did / Checkpoint / Next / Blockers — demanded before every context reset. Every fact in it existed elsewhere first: the current node, attempts and declared skips in work-item-state.json, the coarse phase on the loop:<phase> label, the chronology in the harness's own transcript, the red→green evidence in the commits.

A set of proofs. Eight sections that six review-chain nodes, plus verification's kept gate and the opt-in design-critic-review, read as their exit condition. Those gates exist because they once did not: before issue-167 all six declared sections: and no artifact, so validate-artifacts resolved nothing, returned skipped, and — a skip not being a decision (decision-060) — the chain passed straight through security-review included (decision-063).

The ticket is about the first: "generating the execution log is costing a LOT of tokens — agent harnesses already maintain a log of what has happened, we don't need that to be generated again."

Decision

Delete the narrative. Give each gate one small record of its own, under evidence/.

Node	Record it gates	Section(s)
design-critic-review (opt-in)	evidence/design-critic-review.md	Design critic review
self-review	evidence/self-review.md	Review cycles
critic-review	evidence/critic-review.md	Review cycles
security-review	evidence/security-review.md	Security review (gate)
verification (only when test-planning was declared away)	evidence/verification.md	Verification results
evidence	evidence/final-validation.md	Final validation evidence
capability-docs	evidence/documentation.md	Capability docs, Documentation
reviewer-briefing	evidence/pull-requests.md	Pull requests

And with it:

• The log-entry hook is deleted from the registry and from all 47 entry chains. A node boundary is already an event in the event log and a transition in work-item-state.json.
• The bundled execution-log.md template is deleted; the execution-log role leaves .the-loop/manifest.yaml.
• The multi-repo repos: declaration (issue-183) leaves the log. It shipped in tasks.md's front matter and moved again the same day, on the owner's review, to the phase-selection gate and work-item-state.json — see decision-127, which supersedes this bullet and the alternative below. Absence still means no declaration, never an empty one.
• The checkpoint-then-reset protocol keeps its discipline and loses its prose: tick the checkmarks, commit, keep the label in sync. A fresh window re-enters from work-item-state.json and the first unticked task — which is what "Next:" said, derived rather than written.
• resolve_session's dead-session fallback seeds requirements.md, design.md, tasks.md.

One file per gate, never one shared record. A shared Review cycles section would let critic-review pass on the round self-review wrote — which is what the shared execution log did. Separate files make each assertion its own; a skipped node is routed past before its chain runs, so a skip removes exactly one assertion.

Each gate keeps the validates: verb it already used. produces: is not a synonym for authorship here: it names the artifact a phase is judged by, and the manifest binds each to a phase the parity suite checks in both directions. Five of the six review-chain nodes deliberately carry no phase — the chain was one needs-review label, and minting a label per node would change every consuming repository's vocabulary — so produces: would force either a phase they must not have or a manifest entry P1 cannot satisfy. It also keeps skipped_artifacts (which reads node.produces) empty of these names, so no declared skip can soften a gate the work item still walks.

Existing logs are left exactly where they are. 132 in this repository, more in consuming projects. They are the operator's record of work already done, so they are not in manifest.deprecated (everything there is safe-to-delete plugin internals) and /the-loop:upgrade-the-loop reports them as no longer read rather than removing them — the precedent the relocated learnings tree set. Nothing in the-loop reads one, so a log on disk changes no gate's outcome.

Alternatives considered

Drop the gates with the log. The smallest reading of "remove this functionality": delete the eight sections and the nodes' sections: checks. Rejected — it would also delete validate-artifacts' fail-closed block, parity assertion P5a and decision-063, and return six nodes to reporting success on every run without asserting anything. That is the defect issue-167 was filed for, and it is not what costs tokens: the proofs are a handful of lines written once by the node that ran, while the narrative was a template, 15–17 hook appends and a prose checkpoint per reset. If the gates should go too, that is a separate work item with its own approval.

Keep one smaller shared file. Tidier on disk, and wrong for the reason above — a gate satisfiable by another node's writing. It also costs more tokens than it saves: one file re-read and re-appended at six gates, against six files each written once.

Put repos: in design.md. Defensible (which repositories receive code is a design fact) but one gate further from the PRs being awaited. tasks.md was chosen instead — the artifact the implementation node consumes. Both are superseded by decision-127: the owner's review moved the declaration off artifacts altogether, onto the gate that already freezes every other per-work-item choice.

Consequences

• Breaking for consuming projects: the work-item artifact contract changed. A work item in flight keeps walking — its next gate blocks naming the evidence/<file>.md to write.
• Token cost per work item drops by the template, every log-entry append, and every prose checkpoint; what remains is written once, only for the phases actually walked.
• validates:, onlyWhenSkipped: and the skip vocabulary are reused unchanged. The only test-infrastructure change is _SPEC_FILE in the parity suite, widened so a manifest pathPattern one directory down (evidence/x.md) is matched rather than silently excluded.
• the-loop no longer keeps a chronology of its own anywhere. That is the ticket's premise: the harness already does, and a second copy is generation billed twice.