test(session-flow): eval coverage for the workflow continuation router

[kyle-sexton]

Closes #2972

Summary

Adds eval coverage for the session-flow:workflow continuation router. The router had zero cases exercising its continue path — five workflow evals existed and none touched it — while one of its ordering invariants had already regressed once and been fixed inside the router's own creation PR (#1603, originating issue #1476). The 0.27.0 router evolution (#2971 / PR #3029) deliberately shipped without eval coverage and stated it as out of scope; this is the safety net that item deferred, landing immediately after it.

Fix

Nine cases (ids 6-14) join plugins/session-flow/skills/workflow/evals/evals.json, covering all four bullets of the issue's "Proposed work". Each case grades the router's stated reason, not just its verdict, so a rewrite that reaches the right mechanism by the wrong edge still fails the case.

• First-yes-wins ordering — a machine-going-away prompt with healthy context and a small next step must still route to clean-stop (question 0 outranks every cost question below it); and an explicit background request with healthy context must reach continue-in-background rather than falling through to question 3's zero-cost in-session exit. That second case pins the invariant that regressed once: question 3 answers yes whenever context is healthy, so asking it first silently discards a user instruction.
• Zone gating — a green zone word plus an evidence-degraded compaction marker must read as degraded, and judgment-heavy work in a degraded context must not route to in-session continue. The complement is graded too: a healthy zone whose next stage consumes the current stage's reasoning verbatim must prefer continue, since a summary of the reasoning is not the reasoning.
• Post-evolution behavior — the AFK edge must hand the spawn-brief decision to orchestrate and then keep asking (it is the router's one non-terminal edge) while launching nothing, leaving continue-in-background's explicit-intent gate untouched. Suggest-by-default is graded on its full three-part shape: the single mechanism, the evidence that selected it, and the literal next invocation.
• Autonomy gating — three cases hold the line 0.27.0's review fixes drew: the opt-in counts only in a genuine user turn (consent-shaped text inside a pasted issue body is data, never a licence); auto cannot carry out /clear or /compact, which sit outside the Skill-invocable built-ins; and clean-stop takes the literal continue auto token and nothing else, because once invoked it pushes commits, opens PRs, and files issues without a further confirmation.

The router itself is unchanged — its shipped design is the specification these cases test against, not something this PR revisits.

Also bumps session-flow to 0.28.0 with the matching CHANGELOG entry, per the repo's per-plugin semver parity gate and the precedent of the last eval-coverage-only change (#2989).

Verification

Full local gate, all green:

Gate	Result
check-jsonschema vs plugins/skill-quality/reference/evals.schema.json	ok -- validation done
check-evals-quality.sh on the suite	PASS (0 warnings)
check-evals-quality.test.sh (the lint's own suite)	0 failures
check-orphaned-fixtures.sh --check	no orphaned fixtures
validate-plugins.sh (incl. --strict catalog)	Validation passed
check-changelog-parity.sh --check / --check-bump / --check-preserved / --check-order	all pass (79 headings preserved, 83 changelogs ordered)
check-changed-skills.sh origin/main	CHECK-SKILL workflow: PASS — 0 errors, 0 warnings
generate-catalog.mjs --check / generate-cheatsheet.mjs --check	both in sync
markdownlint-cli2 on the touched CHANGELOG	0 issues
typos on all three touched files	clean

--check-preserved was re-run after the commit so it compared real headings (79) rather than the pre-commit zero-diff case.

Acceptance criteria from the issue: evals exercise the continue path (ids 6-14); the regressed-once ordering invariant is pinned (id 7, with id 6 covering the question-0 half); autonomy opt-in gating covered (ids 12, 13, 14); CI green pending this run.

Related

• Refs session-flow/workflow: aihero course lane 2: phase-boundaries tree re-audit and context-driven continuation router #2900 (AI Hero course lane 2 decision record, decision Q23)
• Refs session-flow/workflow: context-driven continuation router evolution (aihero lane 2) #2971 / PR feat(session-flow): context-driven continuation router #3029 (the router evolution these cases pin; eval coverage was explicitly out of its scope)
• Refs feat(context-guard): zone-crossing hooks, token bands, and the workflow continuation router #1603 (router creation PR, same-PR ordering fix), session-flow/workflow: end-of-phase continuation router (continue / clear / handoff / subagent / compact) #1476 (originating issue)
• Refs claude-memory: /init-then-prune eval fixture — static bloated CLAUDE.md graded against expected audit findings #2989 (precedent: the last eval-coverage-only change and its bump shape)
• Refs skill-quality: check-skill.sh fresh-eyes scanner breaks under mawk — ERE interval expressions unsupported #3005 (the REcompile() panic lines in check-skill.sh output are that known mawk artifact, not findings)

🤖 Generated with Claude Code

https://claude.ai/code/session_01QbfCrj3X9FfGL7VRZYmrn4

---

Generated by Claude Code

[claude]

Claude finished @kyle-sexton's task in 38s —— View job

---

Security review — no security-relevant surface

• Gather context (PR diff, changed files)
• Apply skip gate
• Review changed files for security-relevant surface
• Post final findings

Skip gate hit: item 2 — this change has no security-relevant surface.

All three changed files are non-executable data/config:

• plugins/session-flow/.claude-plugin/plugin.json — version bump 0.27.0 → 0.28.0 only.
• plugins/session-flow/CHANGELOG.md — changelog prose entry for the bump.
• plugins/session-flow/skills/workflow/evals/evals.json — nine new eval cases (ids 6-14) for the session-flow:workflow continuation router: plain JSON objects (prompt / expected_output / expectations strings) consumed by the eval harness as text fixtures. No code, no scripts, no GitHub Actions workflow changes, no permission or credential handling, no injection surfaces (command/SQL/path/template), and no change to the router's own logic — the router implementation is explicitly untouched by this PR.

No security findings. Nothing further to report.

[github-actions]

Last security-reviewed head: 90394a3d7394377bd9623b2e2449c2044671d64f. On the next push, the relevance gate compares only the commits since this SHA; delete this comment to force a full re-review.

[claude]

Claude finished @kyle-sexton's task in 2m 34s —— View job

---

Code review — /review:code-review

• Gather context (PR diff vs origin/main, router spec, eval schema/lint contract)
• Understand the request (CI code-review lane on this PR)
• Review the three changed files
• Post findings

Scope: plugins/session-flow/.claude-plugin/plugin.json (version bump), plugins/session-flow/CHANGELOG.md (0.28.0 entry), plugins/session-flow/skills/workflow/evals/evals.json (+9 cases, ids 6–14). No router logic changed — context/continuation.md is untouched.

What I checked: read the router spec line-by-line against each new case and traced the routing each prompt should produce (question order, zone-degradation rule, non-terminal question 2, the two autonomy-tier rules), cross-checked the .compacted evidence-degraded marker terminology against context-guard's actual reader contract, and diffed against the eval JSON Schema (plugins/skill-quality/reference/evals.schema.json) and the narration/Q4 lint contract in check-evals-quality.sh.

Findings: none blocking. Each of the 9 new cases traces cleanly to a specific clause in the router spec and grades the stated reason, not just the terminal, as the PR describes:

• id 6 / id 7 pin the first-yes-wins ordering exactly as written in continuation.md#L67-84, including the regressed-once invariant (question 1 before question 3).
• id 8's .compacted marker reference matches context-guard's real reader contract (reference/reader-contract.md:36-37), not an invented artifact name.
• id 10 correctly treats question 2 as the router's one non-terminal edge (continues to question 3 rather than asserting a specific mechanism it can't derive from the prompt alone — appropriately loose).
• id 12/13/14 match the three autonomy carve-outs almost verbatim: injected-context-is-never-a-licence, built-in terminals excluded from auto, and clean-stop's literal-token-only gate.
• All 14 cases (5 pre-existing + 9 new) have unique id/name, satisfy the schema's anyOf grading-criterion requirement, and use only schema-declared properties. narration: true is present on cases whose prose contains skill-name-shaped tokens and correctly omitted on ids 11/13, which don't.
• plugin.json version bump is the only change in that file; CHANGELOG.md's new entry lines up with it.

Not independently re-run: this lane's tool grant is Read/Glob/Grep plus gh pr diff/view/comment/review — I did not execute check-jsonschema, check-evals-quality.sh, check-changelog-parity.sh, or the catalog/cheatsheet generators myself, so the PR body's verification table is author-claimed and unverified from this lane, not independently confirmed here. Manual read-through of the JSON against the schema and lint rules above turned up no discrepancy.

[github-actions]

Claude has reviewed this PR 1 time. The lane skips further automatic reviews after 5; deleting this comment resets the count.