release: 0.19.4

[TheAmericanMaker]

Patch release. Completes #211 — which 0.19.3 did not actually finish — plus the /codecarto-next flag discoverability work.

Contents

• fix: block scalars in sequence items, and make /codecarto-next's flags discoverable #216 — block scalars failed inside a sequence (- >-), so a handoff with a decisions: list still died with Invalid YAML indentation. Also makes /codecarto-next's flags self-describing.

This release matters more than its version suggests. 0.19.3 fixed key: >- and not - >-, and the second is the shape an LLM actually writes. On 0.19.3 a phase could run to completion — 25 turns, 73 tool uses, a valid PASS WITH GAPS artifact — and then fail to auto-complete, leaving the pipeline showing 0 complete next to a finished artifact and billed tokens.

Version bumped in all five places tests/release-metadata.test.mjs pins. npm run build clean, 688/688 tests pass.

Surface verification

Run against 0.19.4, driving the packed tarball (the Pi row loads the extension out of the tarball's dist/, not the working tree).

Surface	Run	Result
Pi	/codecarto-init lite then /codecarto-next --auto --llm-steer	full round trip — current_phase: complete, all three lite phases complete, three closeouts, a decision recorded
Claude Code	init → next → validate → complete over stdio	full round trip; see below
Hermes	hermes mcp add then hermes -z	both prompts returned verbatim
Codex	codex mcp add then codex exec	unavailable — blocked, not out of credits this time. See below.

npm pack + node scripts/smoke-mcp.mjs --tarball → 9/9 ok.

The fix, verified with the construct that actually broke

The handoff shape from the real failing run — folded scalars as sequence items:

decisions:
  - >-
    The architecture map treats the two exports as one flat surface rather than
    two subsystems, because neither imports the other…
  - >-
    Module format is recorded as a porting hazard rather than a defect…

codecarto_complete → Marked architecture complete (validation: PASS). Next phase: contracts and 2 decision(s) appended to DECISIONS.md, both folded to single lines (D001, D002). The routed carry_forward — itself a folded description: — landed in status.yaml with target_phase: contracts.

On 0.19.3 that same call returned Invalid YAML indentation near: The architecture map treats….

The user's original workspace was removed before I could re-run against it, so this is a faithful reconstruction of the recorded construct rather than the original file; the original was confirmed parsing against the same fix before it was deleted.

--auto --llm-steer verified, including the steering itself

The Pi round trip used the exact invocation from the bug report. Note that no codecarto-steering messages appear in the headless session log — but neither does codecarto-phase-summary, which certainly fired, so that is a pi -p persistence artifact rather than evidence. The rewriter was therefore exercised directly against the finished workspace:

used=true  skipReason=(none)  prevPhaseId=architecture  promptLen=1428  changed=true

Codex: blocked rather than unverified-for-credits

Credits are back, but codex exec forces approval_policy=never and then refuses every MCP tool call with MCP tool call requires approval, but approval policy is never. -a/--ask-for-approval is not accepted by exec, and -c approval_policy=… / per-server trusted=true do not lift it. The only override is --dangerously-bypass-approvals-and-sandbox, which also removes sandboxing for model-generated shell commands — not something to enable on a maintainer's machine for a verification step.

This is the third release running with Codex unverified, and it is now a known blocker rather than bad luck. Filing separately.

🤖 Generated with Claude Code