v0.6.12

Changed

• Updated @anthropic-ai/claude-agent-sdk from 0.3.287 to 0.3.293 and @openai/codex-sdk from 0.160.0 to 0.161.0. Re-qualified both native adapters with neal compat; no behavior change.

Added

• providers.openai_compatible.structured_output_mode (json_schema or
  json_object, default json_schema) for OpenAI-compatible endpoints that
  support JSON mode but not schema-enforced output, such as direct DeepSeek.
  In json_object mode neal asks for plain JSON, accepts a reply wrapped in a
  ```json fence, and checks it with the same protocol validator. The
  default is unchanged (#81).
• neal usage totals turns, tokens, and cost by provider and role for the
  current run, one run (--run <run-id>), or every run in the repo (--all),
  with --json for scripts. It reads each run's events.ndjson, so it covers
  whole runs, including unfinished ones. RUN_METRICS.json only has whole-run
  totals for finished runs (#82).

Fixed

• OpenAI-compatible runs overstated cost. The AI SDK reports cached input as
  inputTokenDetails.cacheReadTokens, and the cost math never read it, so
  every cached token was billed at the full input rate. Agent loops resend the
  whole conversation each turn, so most input is cached: on an 80%-cached turn
  with cached input priced at a tenth of the full rate, the input cost came out
  about 3.6 times too high. Run metrics now also count
  cache writes and reasoning tokens from the same shape. Costs already recorded
  in old runs aren't recomputed.
• A run could finish even though a manual gate's results showed a failure. A
  gate's checks usually only prove the operator did the work, and once they
  passed, neal forgot the gate. Now the run keeps a record of each gate and
  shows it to the coder and both reviewers, who are told to read the results.
  On a failure, the coder either opens a new gate or splits the plan so the fix
  gets reviewed and the gate runs again (#87).
• When a gate's checks required success and the manual step failed, the run
  was stuck: the checks kept failing, and neal resume --message was refused
  while the gate was open. --message now closes the gate and passes the
  message to the coder (#87).

v0.6.11

Changed

• Updated @anthropic-ai/claude-agent-sdk from 0.3.280 to 0.3.287 and @openai/codex-sdk from 0.156.1 to 0.160.0. Re-qualified both native adapters with neal compat; no behavior change.

v0.6.10

Changed

• Updated @anthropic-ai/claude-agent-sdk from 0.3.274 to 0.3.280 and @openai/codex-sdk from 0.154.0 to 0.156.1. Re-qualified both native adapters with neal compat; no behavior change.

v0.6.9

Changed

• Updated @anthropic-ai/claude-agent-sdk from 0.3.269 to 0.3.274. Re-qualified the native adapter with neal compat; no behavior change.

• A manual gate opens during the coder's first pass on a scope, before any
  review. A scope that both built something for the operator to use (a test
  harness, a procedure, a script) and gated on the operator using it handed over
  unreviewed tooling, and a buggy handoff looped gate, resume, coder fix, gate
  without ever reaching the reviewer. The planner now puts that tooling in its
  own scope ahead of the gate scope, the plan reviewer raises a blocking finding
  on a scope that does both, and a coder that lands in one returns split_plan
  instead of opening the gate. A plan authored one_shot may expand to
  multi_scope for this case only. docs/plan-format.md has a new "Manual
  gates" section and docs/troubleshooting.md covers the reopening gate. The
  plan_author (now 6), plan_reviewer (now 5), and scope_coder (now 4)
  prompt-spec versions bumped (#70).

v0.6.8

Changed

• Updated @anthropic-ai/claude-agent-sdk from 0.3.261 to 0.3.269 and @openai/codex-sdk from 0.153.4 to 0.154.0. Re-qualified both native adapters with neal compat; no behavior change.

v0.6.7

Changed

• Updated @anthropic-ai/claude-agent-sdk from 0.3.260 to 0.3.261 and @openai/codex-sdk from 0.153.2 to 0.153.4. Re-qualified both native adapters with neal compat; no behavior change.

v0.6.6

Fixed

• The read-only reviewer doctrine told every reviewer it had read tools and no
  shell. That's true for Claude and for OpenRouter models, but a Codex reviewer
  reads through a shell in a read-only sandbox, so a model that took the wording
  literally (gpt-6-astra did) had no way to open files and blocked every
  review. The doctrine now says read access may be a read-only shell and that a
  shell is for reading and searching only. The scope_reviewer and
  completion_reviewer prompt-spec versions bumped (#57).

v0.6.5

Changed

• Updated @anthropic-ai/claude-agent-sdk from 0.3.252 to 0.3.260 and @openai/codex-sdk from 0.151.0 to 0.153.2. Re-qualified both native adapters with neal compat; no behavior change.

v0.6.4

Fixed

• A planner block during neal plan no longer loses its reason. The planner's
  explanation is saved with the run and shown on the console, in the run result,
  and in neal status. The run result's next step now says a plain neal resume
  re-runs the planner when that's what the block needs, instead of the generic
  --message hint. A crashed round (a provider error or timeout) now prints its
  error on the console too, instead of a bare "Run failed."; neal status
  already showed it (#49).

v0.6.3

Changed

• Updated @anthropic-ai/claude-agent-sdk from 0.3.251 to 0.3.252. Re-qualified the native adapter with neal compat; no behavior change.