loki-mode 11.17.19 → 11.17.20

package/SKILL.md

@@ -3,7 +3,7 @@ name: loki-mode
 description: An autonomous software factory that knows what it is supposed to deliver, and proves it did. Use when the user says Loki Mode or asks to build, fix or verify software autonomously.
 ---
 
-# Loki Mode v11.17.19
+# Loki Mode v11.17.20
 
 **Current entry points (Loki 10 engine):** `loki "<task>"`, `loki owner/repo#N` and `loki quick "<task>"` run the Loki 10 engine and end with a signed receipt (`loki verify`). `loki backlog owner/repo --all` runs many issues in parallel. Guide: `docs/v10/GUIDE.md`. The skill instructions below, and `loki start`, describe the previous bash loop. PRD-file starts (`loki start ./prd.md`) and the opencode provider still run on it until FC38-SWEEP ports them; the legacy-engine command and the LOKI_ENGINE switch remain removed. Prefer `loki owner/repo#N` for issues.
 
@@ -483,4 +483,4 @@ See `CHANGELOG.md` entries [7.5.7], [7.5.8], [7.5.13] for the per-fix list and r
 
 ---
 
-**v11.17.19 | [Autonomi](https://www.autonomi.dev/) flagship product | core skill**
+**v11.17.20 | [Autonomi](https://www.autonomi.dev/) flagship product | core skill**

package/VERSION

@@ -1 +1 @@
-11.17.19
+11.17.20

package/autonomy/loki

@@ -37563,6 +37563,8 @@ cmd_proof() {
             echo "  md <id>              Paste-able Markdown for a PR comment or Slack"
             echo "  open <id>            Open .loki/proofs/<id>/index.html in a browser"
             echo "  share <id>           Publish the proof page as a GitHub Gist (opt-in)"
+            echo "  bundle <run-id>      Write a portable receipt bundle for offline verification"
+            echo "                       [--out <dir>] (default ./loki-bundle-<run-id>; Bun route)"
             echo ""
             echo "Options for 'share':"
             echo "  --yes                Skip the redaction-preview confirmation prompt"

package/dashboard/__init__.py

@@ -7,6 +7,6 @@ Modules:
     control: Session control helpers (status, atomic writes, origin check)
 """
 
-__version__ = "11.17.19"
+__version__ = "11.17.20"
 
 __all__ = ["__version__"]

package/docs/CLI-REFERENCE.md

@@ -1,6 +1,6 @@
 # CLI reference
 
-Generated from `loki-ts/src/cli/registry.ts` by `scripts/generate-stale-zero.sh` for v11.17.19. Do not edit by hand.
+Generated from `loki-ts/src/cli/registry.ts` by `scripts/generate-stale-zero.sh` for v11.17.20. Do not edit by hand.
 Hidden and legacy-only commands are omitted; see `docs/v10/CLI-MODERN.md` for the full inventory.
 
 | Command | Description |

package/docs/INSTALLATION.md

@@ -2,7 +2,7 @@
 
 Loki Mode, by [Autonomi](https://www.autonomi.dev/), turns a spec or an issue into a verified pull request with a receipt you can check. This page covers installing it, a first run, and the options around it.
 
-**Version:** v11.17.19
+**Version:** v11.17.20
 
 **Engine note:** `loki start` is the primary command. `loki start`, `loki "<task>"`, `loki owner/repo#N` and `loki quick` all run the Loki 10 engine, whatever the flag order or input form (PRD file, issue ref, task or one-word brief). The previous loop runs only on `--engine legacy`, `LOKI_ENGINE=legacy`, or for one call when a flag Loki 10 cannot honor is present. See the [user guide](v10/GUIDE.md) and the [5-minute quickstart](QUICKSTART.md).
 

package/docs/v10/BOARD.md

@@ -1029,7 +1029,7 @@ Source: ~/git/autonomi-dev/research/2026-09-30-adoption/SWARM-PROMPT-ADOPTION.md
 | TEN-COST-06 | xreview claude judge: lean flags and child-ledger line | loki-ts/src/engine10/stages/xreview.ts, loki-ts/tests/engine10/xreview.test.ts | MEDIUM | see Wall in note | released@2026-10-17T03:50Z | TL APPROVE 2e40a4be8, merged for v11.17.14 Shipped in npm 11.17.15 (latest+next) or 11.17.16 (Release run success on e3d445460). |
 | TEN-COST-07 | Project-model discovery on the fast tier, retry on the run model | loki-ts/src/project_model/discover.ts, loki-ts/tests/project_model/project_model.test.ts | MEDIUM | see Wall in note | released@2026-10-17T03:50Z | TL APPROVE e01e37c5e. wave 1; depends: none. Goal: Discovery attempt 1 runs on fastTierModel(env) (runner/model_downgrades.ts:8). It is sonnet unless LOKI_ALLOW_HAIKU, so it never drops below sonnet without opt-in, and it is never weaker than an explicit LOKI_MODEL_OVERRIDE. The existing validation-error retry (attempt 2) runs on the run model, so a weak answer escalates, never ships. The model used is recorded on the project_model result. The schema validation and the owner=provider fallback are unchanged. Wall: Red first, with a stub SessionRunner: - Attempt 1 opts.model equals fastTierModel. - After an invalid answer, attempt 2 opts.model is undefined (run model). - With LOKI_MODEL_OVERRIDE set, both attempts use the override. Mutation: pin attempt 2 to the fast tier; the escalation test goes red. Gate: cd loki-ts && bun test tests/project_model && bunx tsc --noEmit Rules: FAILURE-CLASSES row first for any bug fix (D86); no CHANGELOG edits (Release Manager writes notes); DECISIONS numbers assigned by the Chief of Staff; reuse util/secure_scan.ts, the shared badge helper, keys_cmd export and the TS claude login probe instead of new copies; new CLI flags update autonomy/completions and the help-drift guard. Shipped in npm 11.17.15 (latest+next) or 11.17.16 (Release run success on e3d445460). |
 | TEN-COST-08 | Documented reproducible cost proof run | docs/v10/METRICS.md, docs/EVALUATING.md | LOW | see Wall in note | ready@2026-10-17T01:40Z | wave 4; depends: TEN-COST-04. Goal: docs/v10/METRICS.md gains a 'Cost proof (C-F1)' section that states: - the exact command (attribution.py --execute, arms raw-configured, raw-lean, raw-bare, loki, loki-lean, n=3, pinned model, --max-spend-usd); - the verdict rule from TEN-COST-03; - that results are reported against all three floors, never only the favorable one; - that a run is valid only when the recorded loki version equals VERSION. docs/EVALUATING.md links to it. This is an extension of existing docs, not a new doc. Wall: Red first: grep -n 'Cost proof (C-F1)' docs/v10/METRICS.md exits 1 before the edit. Green: - The documented dry-run command, copied verbatim, exits 0. - grep -n 'raw-lean' docs/EVALUATING.md prints a line. - No em or en dashes: grep -nP '[\x{2013}\x{2014}]' on both files exits 1. Gate: python3 -I benchmarks/bench/attribution.py --dry-run --arms raw-configured,raw-lean,raw-bare,loki,loki-lean && bash tests/test-help-changelog-drift.sh Rules: FAILURE-CLASSES row first for any bug fix (D86); no CHANGELOG edits (Release Manager writes notes); DECISIONS numbers assigned by the Chief of Staff; reuse util/secure_scan.ts, the shared badge helper, keys_cmd export and the TS claude login probe instead of new copies; new CLI flags update autonomy/completions and the help-drift guard. |
-| TEN-QUALITY-01 | Fix brief carries a bounded failure excerpt (assertion, traceback, expected/received) | loki-ts/src/util/check_result.ts (failureExcerpt), loki-ts/src/engine10/stages/verify.ts (runCheck keeps the fail output excerpt; line 315 sample; at most 3 core lines), loki-ts/tests/engine10/check_result.test.ts, loki-ts/tests/engine10/fix.test.ts, loki-ts/tests/engine10/wall_holdout.test.ts | HIGH | see Wall in note | building@2026-10-17T04:27Z | SPLIT per CTO critic into (a) pure failureExcerpt plus fixtures, (b) verify.ts wiring plus holdout test; build as two commits or two agents. wave 2; depends: none. Goal: Each failing non-holdout check's output reaches the fix session as a redacted excerpt of at most 1500 bytes, with at most 6 KB across all groups. The group format is `<rerun cmd>\n<excerpt>`. failureExcerpt() in util/check_result.ts finds the decisive block: pytest `E ` lines plus the last frame, jest/vitest Expected/Received, node AssertionError, go `--- FAIL` with the next lines, and tsc/ruff diagnostic lines. It strips ANSI, makes absolute paths repo-relative and runs util/redact.ts. The signature stays byte-identical, so stall detection at machine.ts:232 and escalation in fix.ts are unchanged. The excerpt lives in a stage-local map inside verify.ts. It is never added to VerifyCheck, never emitted in test.result, and never computed for holdout checks. Wall: Red first: fix.test.ts drives verify then fix on a fixture with a failing pytest `assert add(2,2) == 5`. Today the brief lacks `E assert 4 == 5`, so the test fails. check_result.test.ts covers one fixture each for pytest, jest, node:test, go and tsc. It also asserts the size cap and that an AKIA-style token is redacted. wall_holdout.test.ts asserts that a failing holdout's output text appears in neither failures_grouped, events.jsonl nor the fix brief. Mutations: failureExcerpt returning '' must turn fix.test red; removing the isHold filter at verify.ts:314 must turn wall_holdout red; adding the excerpt to the signature must turn the existing stall test in machine.test.ts red. Gate: cd /Users/lokesh/git/lokimode-anthropic/loki-ts && bun test tests/engine10/check_result.test.ts tests/engine10/fix.test.ts tests/engine10/wall_holdout.test.ts && bun test && npx tsc --noEmit && bun test tests/engine10/budget.test.ts; then bash scripts/local-ci.sh Rules: FAILURE-CLASSES row first for any bug fix (D86); no CHANGELOG edits (Release Manager writes notes); DECISIONS numbers assigned by the Chief of Staff; reuse util/secure_scan.ts, the shared badge helper, keys_cmd export and the TS claude login probe instead of new copies; new CLI flags update autonomy/completions and the help-drift guard. |
+| TEN-QUALITY-01 | Fix brief carries a bounded failure excerpt (assertion, traceback, expected/received) | loki-ts/src/util/check_result.ts (failureExcerpt), loki-ts/src/engine10/stages/verify.ts (runCheck keeps the fail output excerpt; line 315 sample; at most 3 core lines), loki-ts/tests/engine10/check_result.test.ts, loki-ts/tests/engine10/fix.test.ts, loki-ts/tests/engine10/wall_holdout.test.ts | HIGH | see Wall in note | released@2026-10-17T04:35Z | SPLIT per CTO critic into (a) pure failureExcerpt plus fixtures, (b) verify.ts wiring plus holdout test; build as two commits or two agents. wave 2; depends: none. Goal: Each failing non-holdout check's output reaches the fix session as a redacted excerpt of at most 1500 bytes, with at most 6 KB across all groups. The group format is `<rerun cmd>\n<excerpt>`. failureExcerpt() in util/check_result.ts finds the decisive block: pytest `E ` lines plus the last frame, jest/vitest Expected/Received, node AssertionError, go `--- FAIL` with the next lines, and tsc/ruff diagnostic lines. It strips ANSI, makes absolute paths repo-relative and runs util/redact.ts. The signature stays byte-identical, so stall detection at machine.ts:232 and escalation in fix.ts are unchanged. The excerpt lives in a stage-local map inside verify.ts. It is never added to VerifyCheck, never emitted in test.result, and never computed for holdout checks. Wall: Red first: fix.test.ts drives verify then fix on a fixture with a failing pytest `assert add(2,2) == 5`. Today the brief lacks `E assert 4 == 5`, so the test fails. check_result.test.ts covers one fixture each for pytest, jest, node:test, go and tsc. It also asserts the size cap and that an AKIA-style token is redacted. wall_holdout.test.ts asserts that a failing holdout's output text appears in neither failures_grouped, events.jsonl nor the fix brief. Mutations: failureExcerpt returning '' must turn fix.test red; removing the isHold filter at verify.ts:314 must turn wall_holdout red; adding the excerpt to the signature must turn the existing stall test in machine.test.ts red. Gate: cd /Users/lokesh/git/lokimode-anthropic/loki-ts && bun test tests/engine10/check_result.test.ts tests/engine10/fix.test.ts tests/engine10/wall_holdout.test.ts && bun test && npx tsc --noEmit && bun test tests/engine10/budget.test.ts; then bash scripts/local-ci.sh Rules: FAILURE-CLASSES row first for any bug fix (D86); no CHANGELOG edits (Release Manager writes notes); DECISIONS numbers assigned by the Chief of Staff; reuse util/secure_scan.ts, the shared badge helper, keys_cmd export and the TS claude login probe instead of new copies; new CLI flags update autonomy/completions and the help-drift guard. |
 | TEN-QUALITY-02 | Pure static-diagnostic baseline attribution (new vs pre-existing lint) | loki-ts/src/runner/repo_profile.ts, loki-ts/tests/runner/harness_intelligence.test.ts (extends its existing repo_profile describe blocks) | HIGH | see Wall in note | released@2026-10-17T03:50Z | merged 43deeafe9 (HIGH r2 APPROVE 05cafda44) Shipped in npm 11.17.15 (latest+next) or 11.17.16 (Release run success on e3d445460). |
 | TEN-QUALITY-03 | Wire static attribution into verify so pre-existing lint never drives fix rounds | loki-ts/src/engine10/stages/verify.ts (runLintChecks only; at most 4 core lines, helper calls live in repo_profile.ts), loki-ts/tests/engine10/verify_lint.test.ts | HIGH | see Wall in note | released@2026-10-17T04:08Z | v11.17.17; TL APPROVE batch J workflow w9sw9xytw |
 | TEN-QUALITY-04 | Repo map indexes methods and whole large files; cache key versioned | loki-ts/src/engine10/repomap.ts, loki-ts/src/engine10/cache.ts (one-token key rename; net core growth at most 5 lines), loki-ts/tests/engine10/repomap.test.ts, loki-ts/tests/engine10/cache.test.ts | MEDIUM | see Wall in note | released@2026-10-17T03:50Z | TL APPROVE cc1c1b67d. wave 1; depends: none. Goal: repomap.ts records indented Python def and async def under the most recent top-level class as `Class.method`. It raises the symbol-scan cap from 20 KB to 256 KB per file and caps symbols at 300 per file. cache.ts renames the cache file to repomap2-${tree}.json, so a map cached before methods existed is never read. The RepoMap shape is unchanged (symbols are still strings), so isRepoMap and every caller keep working. Wall: Red first: repomap.test.ts writes a 60 KB Python file with `class QuerySet:` and `def bulk_create(self, objs):` past byte 20,000, and asserts the map has `QuerySet.bulk_create`. A file with 1000 defs yields at most 300 symbols. cache.test.ts: a stale repomap-<tree>.json written in the old format is ignored, so a fresh build happens. Mutations: restore the 20 KB cap and the test goes red; restore the old cache filename and the stale test goes red. Gate: cd /Users/lokesh/git/lokimode-anthropic/loki-ts && bun test tests/engine10/repomap.test.ts tests/engine10/cache.test.ts tests/engine10/intake.test.ts tests/engine10/sizing.test.ts && bun test && npx tsc --noEmit && bun test tests/engine10/budget.test.ts; then bash scripts/local-ci.sh Rules: FAILURE-CLASSES row first for any bug fix (D86); no CHANGELOG edits (Release Manager writes notes); DECISIONS numbers assigned by the Chief of Staff; reuse util/secure_scan.ts, the shared badge helper, keys_cmd export and the TS claude login probe instead of new copies; new CLI flags update autonomy/completions and the help-drift guard. Shipped in npm 11.17.15 (latest+next) or 11.17.16 (Release run success on e3d445460). |
@@ -1075,22 +1075,22 @@ Source: ~/git/autonomi-dev/research/2026-09-30-adoption/SWARM-PROMPT-ADOPTION.md
 | TEN-COST-20 | C-F1 seeds ship tests | benchmarks C-F1 seeds | LOW | see note | released@2026-10-17T03:50Z | TL APPROVE 0d1a37eed Shipped in npm 11.17.15 (latest+next) or 11.17.16 (Release run success on e3d445460). |
 | TEN-COST-17 | Raw token classes in the Claude Code ledger | C-F1 harness ledger, test_model_pin.py | HIGH | pytest test_model_pin.py | released@2026-10-17T03:50Z | HIGH r2 APPROVE 232f1fbec Shipped in npm 11.17.15 (latest+next) or 11.17.16 (Release run success on e3d445460). |
 | TEN-RELEASE-INSTANT | Every release publishes npm latest and next immediately (D102) | .github/workflows/release.yml, promote.yml removed | HIGH | see note | released@2026-10-17T03:50Z | HIGH r3 APPROVE 7e0e38c77; D98 round-limit exception recorded under D99 Shipped in npm 11.17.15 (latest+next) or 11.17.16 (Release run success on e3d445460). |
-| TEN-GATES-05b | Spec-derived auth probe: negation window fixes | autonomy/lib/functional-verify.py, tests/test_functional_verify.py | HIGH | pytest tests/test_functional_verify.py | building@2026-10-17T04:31Z | HIGH r2 BLOCK aaf27e569: trailing-slash redirect false green, curly apostrophe and article negation false reds; r3 under D99 |
+| TEN-GATES-05b | Spec-derived auth probe: negation window fixes | autonomy/lib/functional-verify.py, tests/test_functional_verify.py | HIGH | pytest tests/test_functional_verify.py | building@2026-10-17T04:35Z | HIGH r2 BLOCK aaf27e569: trailing-slash redirect false green, curly apostrophe and article negation false reds; r3 under D99 |
 | TEN-RELEASE-INSTANT-F1 | Empty CUR from a failed npm view fails closed; rename the stale publish step | .github/workflows/release.yml | LOW | see note | released@2026-10-17T03:50Z | TL APPROVE 29ca496f1, ships v11.17.16 Shipped in npm 11.17.15 (latest+next) or 11.17.16 (Release run success on e3d445460). |
 | TEN-COST-18-F1 | cost_claim: a non-object JSON row is skipped, not an AttributeError | C-F1 harness cost_claim | LOW | see note | released@2026-10-17T03:50Z | TL APPROVE 924cec1c8, ships v11.17.16 Shipped in npm 11.17.15 (latest+next) or 11.17.16 (Release run success on e3d445460). |
 | TEN-GROUP-AUTH-F1 | kpis iteration_count does not count group cost records | loki-ts kpis | LOW | see note | released@2026-10-17T04:31Z | r1 BLOCK 1c59103ad (writer kind unasserted); r2 adds writer-reader test and stats.ts sibling |
-| TEN-COST-09 | Skip a Wall that cannot run | engine10 wall stage | HIGH | see note | review@2026-10-17T04:31Z | architect cost plan |
-| TEN-COST-10 | Plan on haiku for small and normal tasks | engine10 plan stage | MEDIUM | see note | building@2026-10-17T04:31Z | architect cost plan |
+| TEN-COST-09 | Skip a Wall that cannot run | engine10 wall stage | HIGH | see note | building@2026-10-17T04:35Z | architect cost plan |
+| TEN-COST-10 | Plan on haiku for small and normal tasks | engine10 plan stage | MEDIUM | see note | rejected@2026-10-17T04:35Z | D99: FC-35 pins the router-on plan to opus and router-off must stay byte-identical, so the haiku plan pin was dead code (HIGH BLOCK bc80a88b6); quality outranks cost. architect cost plan |
 | TEN-COST-11 | Wall on haiku, re-authored on sonnet on failure | engine10 wall stage | HIGH | see note | ready@2026-10-17T03:30Z | architect cost plan |
-| TEN-COST-12 | 5-minute cache TTL by default | runner cache settings | MEDIUM | see note | building@2026-10-17T04:31Z | architect cost plan |
+| TEN-COST-12 | 5-minute cache TTL by default | runner cache settings | MEDIUM | see note | released@2026-10-17T04:35Z | architect cost plan |
 | TEN-COST-13 | Lean prefix by default | runner prompt prefix | MEDIUM | see note | released@2026-10-17T04:31Z | architect cost plan |
 | TEN-COST-14 | Wall brief output budget | engine10 wall brief | LOW | see note | released@2026-10-17T04:24Z | architect cost plan |
-| TEN-COST-15 | Plan output trim | engine10 plan stage | LOW | see note | ready@2026-10-17T03:30Z | architect cost plan |
+| TEN-COST-15 | Plan output trim | engine10 plan stage | LOW | see note | parked@2026-10-17T04:35Z | parked with TEN-COST-10 (rejected). architect cost plan |
 | TEN-COST-16 | Implement on haiku first (opt-in) | engine10 implement stage | HIGH | see note | building@2026-10-17T04:15Z | architect cost plan |
 | TEN-COST-19 | Group cost in loki rows of C-F1 | C-F1 harness | LOW | see note | ready@2026-10-17T03:30Z | after COST-04 |
 | TEN-COST-21 | C-F1 .venv and keep_evidence | C-F1 harness | LOW | see note | ready@2026-10-17T03:30Z | after COST-04 and COST-20 |
 | TEN-COST-22 | Cache read pricing correction | model-pricing.json | LOW | see note | rejected@2026-10-17T03:30Z | premise false: Sonnet 5.5 cache read is $0.10 (0.05x) already |
-| TEN-COST-23 | Investigate multi-unit launcher exit 1 | features/speed launcher | HIGH | see note | building@2026-10-17T04:31Z | architect cost plan |
+| TEN-COST-23 | Investigate multi-unit launcher exit 1 | features/speed launcher | HIGH | see note | released@2026-10-17T04:35Z | architect cost plan |
 | TEN-COST-24 | Opt-in Batch API path for single-shot judge calls under an API key | judge call sites | MEDIUM | see note | building@2026-10-17T04:15Z | never used in loki start; subscription runs gain nothing from batch |
 | TEN-GATES-03-F1 | verify --require-exogenous fails closed with --allow-unsigned and ignores zero-test checks | loki-ts/src/engine10/verify_cmd.ts; loki-ts/tests/engine10/verify_cmd.test.ts | HIGH | an UNSIGNED agent-only receipt with --require-exogenous --allow-unsigned exits 4; an exogenous pass with n=0 does not count toward k | building@2026-10-17T04:15Z | from GATES-03 HIGH review non-blocking findings 1 and 2; batch O |
 | TEN-ADOPTION-03-F1 | proof share scan catches tokens (ghp_, AKIA, sk-ant) and proof registry lists badge, md, bundle | autonomy/lib/secure-scan.py or loki-ts/src/util/secure_scan.ts; loki-ts/src/commands/proof.ts; loki-ts/src/cli/registry.ts; tests | MEDIUM | a bundle with a ghp_ token in events.jsonl shows a finding in the share preview; registry proof subcommands include badge, md, bundle | building@2026-10-17T04:15Z | from ADOPTION-03 TL review notes 1 and 2; batch O |

package/docs/v10/CLI-MODERN.md

@@ -172,7 +172,7 @@ Classes: KEEP-MODERN (v10-native or engine-neutral), UPDATE (live but needs mode
 - enterprise: 2
 - explain: 5
 - export: 6
-- github: 5
+- github: 6
 - import: 1
 - init: 21
 - issue: 2

package/docs/v10/FAILURE-CLASSES.md

@@ -997,3 +997,10 @@ L0 review rule (ENGINE-LAWS.md L0): a fix that adds an `if` or a regex about the
 - Siblings: deep.ts was the only engine10 stage importing runner/quality_gates.ts (git grep); xreview.ts already imports providers/claude_flags.ts. pack-npm also lacked a loki-ts install, fixed in 386e73f32.
 - Mechanism: REVIEW_GUARD_DENYLIST moves to providers/claude_flags.ts; quality_gates.ts re-exports it; deep.ts imports the leaf.
 - Fixture: tests/test-control-plane.sh builds serve.ts with a metafile and fails if runner/sdk_invoker appears (red with the old import: 3 hits; green: 0).
+
+## FC-TEN-COST-23 Decomposer emits a unit with an empty write set and unit mode refuses it (EMPTY-WRITESET-UNIT)
+- User-visible vs raw: every multi-unit run of coord-1-api-web exited 1 with "unit launcher exit 1; 2 unit(s) did not pass" after about 50s, the spurious unit having failed with exit 2 in 1s and $0; the raw reason ("unit mode was requested but its spec file is invalid") never reached the user.
+- Law broken: a decomposer output must be runnable by the stage that consumes it; unit_mode.ts parseText rejects writeSet.length === 0 and the engine refuses an unfenced run, so a file-less prose bullet ("`limit` must be an int ...") became a unit that could never start.
+- Siblings: decompose.ts items with no matching file (also shared-file-only items); fix_units.ts units (always carry a write set); the u1 PARTIAL exit 1 in the same evidence is the separate no-test-runner seed cause (TEN-COST-20, TEN-COST-21), not this class.
+- Mechanism: decompose() folds every item that names no write-set file into the first unit that owns files; if no item names a file all items form one unit and route.ts falls back to the sequential run (fewer than 2 units).
+- Fixture: tests/engine10/decompose.test.ts "no unit has an empty write set" and "a spec where no item names a file collapses to one unit"; repro: LOKI_UNIT_SPEC with writeSet [] and bin/loki start gives rc=2 "spec file is invalid".

package/docs/v10/PROGRESS.md

@@ -2427,3 +2427,6 @@ The swarm is stopped: .loki/V10-STOP exists, the loop has no wake-ups, and the c
 
 ## 2026-10-17T04:31Z (CoS) v11.17.19: COST-13 and GROUP-AUTH-F1
 - v11.17.17 publish-npm success (Release run on 1fa2b21cc; npm-visible job pending on propagation). v11.17.18 (COST-14) held behind that run, now pushed together with v11.17.19. TEN-COST-13 TL APPROVE 0b15dc2ba (session 225 pass, runner 1293 pass, 2 mutations caught). TEN-GROUP-AUTH-F1 r2 TL APPROVE 367b4e042 (writer, stats and shared predicate mutations each fail a test). Batch L blocks: GATES-05b r3 BLOCK 7eef05bcb (base_url prefix false green, regex tail unguarded); COST-12 BLOCK b96ffb5f9 (call-site wiring untested); COST-10 BLOCK a0bd42473 (router-off goldens red); COST-09 reviewer returned no verdict; COST-23 build failed on the fixed USER/LOGNAME env drop. r2 rounds dispatched. TEN-QUALITY-01 building; TEN-STABILITY-03 parked under D102.
+
+## 2026-10-17T04:35Z (CoS) v11.17.20: P0 CI fix, QUALITY-01, COST-12, COST-23
+- v11.17.19 Release run success (npm next 11.17.19, latest 11.17.17 propagating). P0-CI-RED TL APPROVE ec63f4398 (6 mutations each caught). TEN-QUALITY-01 r2 HIGH APPROVE 80dd9af12 (total cap byte-exact 6144 on CJK, signature stable, holdout and redaction mutations caught). TEN-COST-12 r2 TL APPROVE e177b6cd2 (both call sites pinned). TEN-COST-23 TL APPROVE 8c0bbaaf8. Decisions under D99: TEN-COST-10 rejected (router-on plan stays pinned to opus by FC-35, router-off stays byte-identical, so the haiku plan pin was dead code; quality outranks cost); TEN-COST-15 parked with it. TEN-COST-09 r2: skip the Wall only when holdout is off (r1 HIGH BLOCK: skipping discarded the holdout run). TEN-GATES-05b r5: resolve relative Location headers before comparing.