Skip to content

Commit 598fb27

Browse files
jmaglyjmagly
committed
chore(release): cut v2026.5.6 — agent-loop completion inference + auto-compact discipline
Version: 2026.5.5 -> 2026.5.6 (CalVer, no leading zeros). Lockstep: package.json + .claude-plugin/marketplace.json metadata.version. CHANGELOG.md and docs/releases/v2026.5.6-announcement.md cover: - New infer-completion-criteria skill (5-layer pipeline: task verb, context files, manifests, CI config, .aiwg artifacts) that turns "agent-loop fix the failing tests" without --completion into a measurable, project-grounded verification command. - New auto-compact-continue HIGH rule (deployed in v2026.5.5's incident-fix commit; surfaced here for visibility now that the agent-loop changes pair with it). Codifies the "should I keep working = always YES" discipline backed by Anthropic's Effective Harnesses for Long-Running Agents (REF-909) and Compaction docs (REF-910). - CI workflow ordering fix in .gitea/workflows/npm-publish.yml (build before test, matching ci.yml), closing the v2026.5.5 publish regression mode. UAT: npm run uat -> 93 passed across 5 test files.
1 parent d28ec20 commit 598fb27

4 files changed

Lines changed: 325 additions & 2 deletions

File tree

‎.claude-plugin/marketplace.json‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -6,7 +6,7 @@
66
},
77
"metadata": {
88
"description": "AIWG - Modular agentic framework for SDLC, marketing automation, and workflow orchestration. 90+ agents, 95+ commands, 30+ skills.",
9-
"version": "2026.5.5",
9+
"version": "2026.5.6",
1010
"homepage": "https://aiwg.io",
1111
"repository": "https://github.com/jmagly/aiwg",
1212
"documentation": "https://docs.aiwg.io",

‎CHANGELOG.md‎

Lines changed: 51 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -9,6 +9,57 @@ and this project uses [Calendar Versioning (CalVer)](https://calver.org/) with n
99

1010
_Nothing yet for the next release line._
1111

12+
## [2026.5.6] - 2026-05-14 — "Agent-loop completion inference + auto-compact discipline"
13+
14+
Two behavioral upgrades to make long-running and iterative work survive context pressure and reach measurable completion without operator hand-holding, plus a CI workflow ordering fix that prevents v2026.5.5-style stable-publish regressions.
15+
16+
### Why this matters to users
17+
18+
| What changed | What it gives you |
19+
|---|---|
20+
| **`agent-loop` infers `--completion` from project state** | Run `agent-loop "fix the failing tests"` without `--completion`. A new `infer-completion-criteria` skill walks five evidence layers — task verb → `CLAUDE.md`/`AGENTS.md`/`AIWG.md` Development sections → package manifests (`package.json`, `Cargo.toml`, `pyproject.toml`, `go.mod`, `pom.xml`, …) → CI config (`.github/workflows/`, `.gitea/workflows/`, GitLab/CircleCI/Jenkins) → `.aiwg/` artifacts (test-strategy, related use cases by ID match, prior progress files) — and proposes a measurable criterion with rationale and confidence level. High-confidence proposals auto-adopt with `--auto-criteria`; otherwise the proposal is surfaced for confirm/edit/abort. Refusal path is explicit: tasks like "make it better" with no derivable measurable gate get a refusal with concrete rephrasing suggestions. |
21+
| **New `auto-compact-continue` rule (HIGH)** | The answer to "should I keep working?" is always YES — until measurable completion criteria are met or the user redirects. Context pressure, long tool output, and crossing iteration N are not scope questions. The rule codifies the auto-compact-and-continue discipline backed by Anthropic's [Effective Harnesses for Long-Running Agents](https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents) and the [Compaction docs](https://platform.claude.com/docs/en/build-with-claude/compaction): write progress files at `.aiwg/working/<task>-progress.md` every 10–15 tool calls (REF-122 aggressive vs passive: 22.7% vs 6% savings), let the platform auto-compact, recover via the AIWG durable substrate (activity log, progress file, git, memory, `CLAUDE.md`/`AGENTS.md`/`AIWG.md`). |
22+
| **CI: build before tests in `npm-publish.yml`** | The v2026.5.5 stable publish failed because tests ran before build, and several tests assert on `dist/` output (regression test for #1001). The Gitea workflow now matches `ci.yml`: typecheck → build → test → publish. Without this, the next stable publish would have hit the same failure mode. |
23+
24+
### Added
25+
26+
- `agentic/code/addons/agent-loop/skills/infer-completion-criteria/` — new skill (~280 lines). Deterministic 5-layer inference pipeline with structured YAML output (criterion, verification command, rationale, confidence, alternatives, max-iterations suggestion). Documents edge cases for monorepos, multi-language projects, broken-test suites, use-case acceptance criteria, and the refusal path. Wired into `agent-loop/manifest.json` skills array.
27+
- `agentic/code/addons/aiwg-utils/rules/auto-compact-continue.md` — new HIGH rule (~250 lines). Codifies the auto-compact-and-continue discipline with 8 mandatory rules, interaction with `vague-discretion`/`anti-laziness`/`human-authorization`/`instruction-comprehension`/`skill-discovery`/`activity-log`/`context-budget`, recovery protocol after compaction, named exceptions, and platform applicability across all 10 supported providers. (Landed in v2026.5.5's incident-fix commit f87ba48c; called out in this release for visibility since the user-facing impact is felt now that the agent-loop changes pair with it.)
28+
- Two new research references in companion `research-papers` repo:
29+
- **REF-909** — Anthropic, *Effective Harnesses for Long-Running Agents* (Nov 2025). Initializer-agent / coding-agent pattern, `claude-progress.txt`, "failed approaches" as load-bearing artifact across context resets.
30+
- **REF-910** — Anthropic, *Compaction* (Claude API Documentation, 2026). Auto-compact mechanics, `## Compact Instructions`, what survives compaction.
31+
32+
### Changed
33+
34+
- `agentic/code/addons/agent-loop/skills/agent-loop/SKILL.md` — completion-inference section now delegates to the new `infer-completion-criteria` skill. The previous inline 7-row Node-centric pattern table is demoted to a "last-resort fallback" with explanatory note. Related-skills list adds `infer-completion-criteria` and `agent-loop-ext`.
35+
- `agentic/code/addons/agent-loop/skills/ralph/SKILL.md` — `--completion` is now optional. Phase 1 initialization invokes inference when omitted; high-confidence proposals or `--auto-criteria` adopt silently, otherwise the user is asked via `AskUserQuestion` (per `native-ux-tools`). Loop also writes the resolved criterion and rationale into the loop's progress file per `auto-compact-continue`. New flag: `--no-infer-completion` for explicit-required behavior.
36+
- `agentic/code/addons/agent-loop/skills/agent-loop-ext/SKILL.md` — same `--completion`-optional treatment for the crash-resilient external loop, with TTY-aware confirmation (interactive: confirm; headless/CI: adopt high-confidence proposals, fail fast with diagnostic on low confidence). Proposal persisted to `.aiwg/ralph-external/<run-id>/inferred-completion.yaml` for crash recovery.
37+
- `agentic/code/addons/agent-loop/agents/ralph-verifier.md` — gained a "Companion skill" section documenting that when the criterion was inferred, the verifier can reference the rationale chain from the progress file / inferred-completion.yaml in its verification reports.
38+
- `agentic/code/addons/agent-loop/manifest.json` — added `infer-completion-criteria` to the skills array.
39+
- `agentic/code/addons/aiwg-utils/manifest.json` — added `auto-compact-continue` to the rules array; `RULES-INDEX.md` updated to 21 rules with the new HIGH-tier entry placed at the top of the index.
40+
- `.gitea/workflows/npm-publish.yml` — Build step now runs **before** Run tests, matching `ci.yml`. Several tests assert on `dist/` output (e.g. `test/unit/cli/validate-metadata-import.test.ts` asserts the import path resolves to `dist/src/plugin/metadata-validator.js`); pre-2026.5.6 the workflow ran tests against stale build output and the stable publish at v2026.5.5 hit this. Inline comment cites the v2026.5.5 incident so future contributors understand the ordering constraint.
41+
42+
### Fixed
43+
44+
- Stable npm publish path: tests now run against current `dist/` artifacts. The v2026.5.5 publish failure mode (test references resolving against stale or missing build output) is closed.
45+
46+
### Migration notes
47+
48+
No breaking changes. Existing `agent-loop` / `ralph` / `agent-loop-ext` invocations with explicit `--completion` continue to work unchanged. The new behavior only activates when `--completion` is omitted.
49+
50+
If you want the old hard-error-on-missing-completion behavior, pass `--no-infer-completion`. If you want CI-style fully-automated runs that adopt the inferred criterion without confirmation, pass `--auto-criteria`.
51+
52+
### Companion: research-papers corpus
53+
54+
The AIWG research corpus at `git@git.integrolabs.net:roctinam/research-papers.git` gained REF-909 and REF-910 as the load-bearing citations for the new `auto-compact-continue` rule. Both are GRADE LOW (vendor documentation) but authoritative for Claude-specific patterns and reflect production experience from the Claude Code team.
55+
56+
### Links
57+
58+
- [Effective harnesses for long-running agents](https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents) — Anthropic Engineering, Nov 2025
59+
- [Compaction — Claude API Docs](https://platform.claude.com/docs/en/build-with-claude/compaction) — Anthropic, 2026
60+
- REF-122 *Active Context Compression* (Verma, 2026) — already in corpus; the empirical basis for aggressive in-session compression
61+
- Release announcement: [docs/releases/v2026.5.6-announcement.md](docs/releases/v2026.5.6-announcement.md)
62+
1263
## [2026.5.5] - 2026-05-14 — "Cross-provider discover-first parity"
1364

1465
Two user-facing changes plus the deployment-pipeline fix that makes the

0 commit comments

Comments
 (0)