v2.7.0 — Diff-scoped story loop and harness guards

Half the subagent dispatches for the same guarantees, and a set of guards that close the ways a
pipeline run went wrong while it was being built. The first half is a measured change to the story
loop; the second is what running that loop on real work taught.

The study: where does a delivery's cost go?

Question. A delivery spends tokens in two places: the guides each dispatch loads, and the number
of dispatches. v2.6.0 measured the first and cut it 3.5%; the dispatch graph was untouched at 40 per
five-story delivery. Which of the two is worth changing?

Method. scripts/bench-context.py models the dispatch graph and the guide tokens each path
loads, run against the v2.6.0 tree and the new one. Per-agent costs come from real dispatches, not
estimates: every agent run in this release was logged with its token count.

Result.

Per five-story delivery	v2.6.0	2.7.0	Change
Subagent dispatches (uniform roster)	40	20	−50%
Subagent dispatches (mixed rosters)	—	18
Guide tokens loaded	244,170	216,309	−11.4%

The mechanical sensors read far less than an agent would: a security scan prints 365 tokens where
the changed files are 66k (182×), a tautology scan 65 where a 38-file test suite is 52k (797×), the
duplication gate 91 where raw jscpd output is 6.8k (75×).

Why it moved. Dispatches fell because work that was agent judgment became arithmetic or a
sensor: the verdict is verdict.sh, not an agent; story generation for a one-row manifest is a
template fill; the story-level PR review read the same diff the Reviewer had already read; Stress
runs only when the diff warrants it. Validation reads the story's diff, not the tree.

What the model does not show. Fix rounds. The estimate counts first passes; on stories that ran
the full roster, rework added 40–60% on top. The guide-token saving also shrank from −14.2% to
−11.4% once the rules below were written into every agent's guide — a deliberate trade of tokens
for guarantees, measured rather than assumed.

Added

• Rosters. Each story declares a roster — cosmetic, light, standard, full — and it
  decides which agents run. Mechanical gates run on every roster without exception.
  classify-diff.sh checks the declaration against the real diff and only ever escalates: a file
  that looks like style but holds code, a rename that disguises code as style, a symlink, or a
  deleted source file cannot lower the roster.
• One-command mechanical checkpoint. checkpoint-m.sh runs every mechanical check a story must
  pass and reports every failure at once: spec rows by exact name, the exact changed-file set
  (renames count both paths, deletions count), no commits beyond the base, no leftover
  falsification markers, the roster, the repository's gate command and duplication. A sub-check
  that cannot measure fails closed. No agent validates code this checkpoint has failed.
• Self-restoring falsification. break-run.sh applies one literal mutation — refused unless it
  matches exactly once — requires the named test to fail by its own name, and restores the file
  byte for byte on every exit it can trap, keeping the pristine copy if a restore ever fails. It
  refuses symlinks and paths outside the repository and locks the file against concurrent runs.
• Release-content guard. Session memory, handoff notes and generated test output are rejected
  from commits and pushes by path, from patterns in a committed .devkit/release-exclude. Every
  published commit is checked, merge commits included; a git failure blocks instead of passing.
• Execution-options checkpoint. After a plan is approved, bench-context.py --manifest shows
  the operator dispatches and a token range per option — with first-pass and with-rework figures,
  each cost labelled measured or assumed — and no code starts until one is chosen.
• Oversized-prompt warning. dispatch-budget.sh warns when a single agent prompt is too long or
  names too many files; oversized prompts were the cause of every agent stall observed.

Changed

• The story loop validates the story's own diff, dispatches by roster, runs the Reviewer and Stress
  in parallel after QA, and re-checks narrowly after a fix. The verdict is verdict.sh.
• Agent guides carry scope rules: edit only the listed files, never revert work you did not make,
  probe only under $TMPDIR, falsify through break-run.sh, make every error path fail closed and
  test it. An agent's report is a claim; the evidence is what the checkpoint reproduces.
• PROGRESS.md is local session memory: git-ignored and never committed.

Lessons behind the guards

Each failure class below was observed while building this release; each has a guard or a rule.

Failure	Guard
An agent reported work or evidence that did not exist	checkpoint-m.sh checks rows by name and the exact file set; the orchestrator reproduces evidence with break-run.sh
A test failed without naming itself, or passed for the wrong reason	break-run.sh accepts only a failure that names the row
An agent died mid-test and left code deliberately broken	Break markers; the checkpoint fails while one exists
A mutation hit a comment instead of code and proved nothing	break-run.sh refuses a pattern that is not unique
An agent committed, reverted or created files outside its scope	Exact file-set and commit checks; scope rules in every guide
A check that could not measure reported a pass	Every sensor exits 2 (unmeasured) on a failure it cannot see past
A classifier trusted a file's name over its content	Content is scanned before a roster is ever lowered
Oversized prompts stalled agents	The prompt-size warning; split large dispatches
Estimates ignored rework	Rework factors in the execution-options table
A timing bound held only on an idle machine	Test waits are sized for a loaded machine
Session notes shipped in a release	The release-content guard

Still open: an agent told to work in a worktree can edit the main checkout instead. A guard that
refuses edits outside the active worktree is the next change.

Versions

Plugin	Version	Why
coding-pipeline	2.6.0 → 2.7.0	Rosters, new verify scripts and guards (features).
devtools	1.3.0 → 1.3.1	handoff docs: PROGRESS.md is git-ignored, never committed.
engineering	1.5.1	Unchanged.
pr-workflow	1.5.1	Unchanged.