Handoff: what to do on 1 October 2026 and in which order

[martinfrancois]

Handoff: prove PR #94, release 1.2.2 to the registry

State on 2026-09-21: the registry serves 1.2.0 (quality 100, 2.22x). The v1.2.1 GitHub release exists but its Tessl publish failed at the gate in scripts/validate_publish_ready.sh (tessl review run --threshold 100), because the released SKILL.md scores 91 under the current Tessl reviewer (#88, run 35561430400). PR #94 rewrites SKILL.md to score 100 (review run 01a0c22c-f9e8-7376-933c-240b76d3cf1d) without changing any rule; it has no hosted evidence yet. The maintainer decided: keep 1.2.1 GitHub-only, publish 1.2.2 once #94 is proven. Do this after java-functional-style-skill PR #18 and its 1.0.0 release, and after merging the optionals release PR, because this is the most expensive proof.

Steps

1. git fetch origin && git checkout fix/skill-review-100, rebase onto main if behind, force-push with lease. Run the local validators from docs/agents/workflow.md.
2. Quality review must still be 100 on the rebased head:
   tessl review run --workspace martinfrancois --threshold 100 --force skills/java-streams/SKILL.md
3. scripts/pre_submit_gate.sh --plan-only prints the targeted probe order for this runtime change. Run the probes it names, one or two at a time, with scripts/run_eval_suite.sh <suite> <scenario> -- --wait -y. Every with-context result must be 100 before widening.
4. Full main suite, both variants (80 credits for 4 scenarios): scripts/run_eval_suite.sh main -- --wait -y. The ratio must stay at or above 2.22x with all four at 100 with context.
5. Reference, both variants per docs/agents/evals.md (120 for 6 scenarios), or with context only (60) if the budget is tight; state which in the PR: scripts/run_eval_suite.sh reference -- --wait -y.
6. Regression with context only (190 for 19 scenarios): scripts/run_eval_suite.sh regression -- --wait -y. All at 100.
7. Composition with the companion, from a checkout of java-functional-style-skill next to this one: scripts/run_composed_eval.sh ../java-streams-skill main -- --wait -y (40). All at 100.
8. Record run IDs in the numbering notes and criteria-meta.json, fill the "Hosted Evidence" table in the PR body, push.
9. Squash-merge fix(skill): tighten the stream workflow for the current Tessl reviewer #94 (title fix(skill): ...). Release Please opens "chore(main): release 1.2.2"; merge it. This time the publish gate passes (review 100) and the registry evaluates 1.2.2.
10. Confirm on https://tessl.io/registry/martinfrancois/java-streams: version 1.2.2, quality 100, score at or above 2.22x. If the registry run drops below 2.22x, read tessl eval view <id> --json for the scenario that moved before changing anything; variance on one scenario is normal, a with-context miss is not.

Budget: about 400 credits with both-variant reference, about 340 with context-only reference. Together with the functional-style proof (about 300) and two registry evals this fits one monthly window only if nothing is repeated, so do the functional-style repository first and check tessl org usage before each stage.

Also open, lower priority: #44 (identity-mapper eval, now solved by the default solver; comment on the issue explains the re-scope), #93 (share the commitlint install block).

Ground rules for whoever picks this up

• Do this on or after 1 October 2026, when the Tessl free plan's credit window resets (1,000 credits per month, 300 eval credits per day). Check first: tessl org usage.
• Sign in with tessl login --device (the loopback flow prints nothing on a headless host). Any machine works; clone all three repositories side by side, because the composition runner in java-functional-style-skill reads the sibling checkouts by relative path (../java-streams-skill, ../java-optionals-skill).
• Costs, so the budget can be planned: one scenario with both variants costs 20 credits, with context only 10, a quality review 10, a registry eval at publish time 30 to 60. Cells drop at random on this plan; tessl eval retry <run-id> creates a new run that reuses the completed cells. Never rerun a run because it looks slow; poll it with tessl eval view <run-id> --json.
• Every hosted result must be recorded with its run ID in the numbering notes (evals/NUMBERING.md, evals-reference/NUMBERING.md, evals-regression/NUMBERING.md) and the per-scenario criteria-meta.json, never in docs/agents/* (the validator rejects run IDs there).
• Any commit that touches the runtime bundle (skills/, and rules/ where it exists) invalidates every with-context result; freeze the runtime text before spending credits.
• Commitlint rejects merge commits in these repositories, and the streams and optionals rulesets require branches to be up to date, so update a branch by rebase and force-push with --force-with-lease --force-if-includes (the maintainer approved that for these branches), never with a merge commit.
• Model selection (--model) is a paid-plan feature and returns an entitlement error on the free plan; use the default solver.

AI Assistance (if used)

• AI-assisted issue
• I confirm I understand and reviewed this request