feat(orchestrator): delegate for a reason, not for size - #1794

…d of task size

…fication module

…ation pay for themselves

…nd contracts

…, and gate Close on pending tasks

… per unit

…exploration

Define read-only for gentle-ai-verify as no edits to the repository or its
git state, and let it run the spec's examples and probes on its first launch
inside a fresh mktemp scratch copy with the project's own commands. The
verify handoff names probe commands and the scratch location instead of a
bare read-only. The ODD Explore step no longer delegates exploration of files
the parent will read anyway.

… results

pi --mode json runs the same one-shot print runner as pi -p and disposes the
runtime after the prompt, so background subagents launched there were killed
and the turn ended empty. Both modes now default to task mode and reject
explicit background launches. Task-mode subagent results now start with the
completion header that names the task id and label, so the model can continue
a task without guessing ordinal ids.

Keep the per-unit independent verify on real high risk and for
smaller-model writer code, but when the orchestrator itself wrote the code
inline across several deliveries of one feature, run one verify at the end
of the feature. Raise the lazy verification module budget to 6,500 bytes.

The first-launch probe rule did not say how to build the scratch copy, and
natural choices such as git worktree, a symlinked node_modules, or project
commands writing to HOME could write outside it. Copy the workspace files
without .git, never use worktrees, stashes, or symlinks into the workspace,
and keep HOME and TMPDIR inside the scratch directory.

The Evidence-budget rule fired after a couple of reads on small tasks the
parent then implemented inline, adding a delegated exploration that the
parent re-read anyway. Reading to prepare an inline write never fires it.

Verify reported behavior already present at the baseline and values far
outside the realistic domain as blockers, forcing needless correction
rounds. A blocker now needs a change-caused defect or unmet spec item
reproduced with realistic input; baseline behavior not asked to change and
out-of-domain values are advisories. Inputs the program accepts, including
hand-edited files, are realistic; rules for one entry path hold for all;
silently ignoring an explicit option and unrequested changes to existing
command output stay blockers.

…check

Nothing capped verify/fix/recheck cycles, and each full recheck surfaced new
findings, so one delivery could loop until the cell timed out. Blockers now
get one correction batch and one re-verify limited to those blockers; any
still open stop as a single decision for the user, and advisories never
start a correction. Raise the lazy verification budget to 6,800 bytes.

A recheck that shows the same blocker still failing earns one more
automatic correction; a new finding never does. The verifier also keeps
its scratch path out of files outside the scratch and disables the npm
update notifier during probes. Raise the lazy verification budget to
7,000 bytes.

Partial, blocked, unavailable, or exhausted proof now ends in one Needs
your decision result instead of more verification, and the close gate
accepts that result as a valid stop. Writer self-review and worker
validation get one correction per failing check, a second only when the
same check still fails. A quick check runs once, and an unavailable
verifier is reported instead of retried. Raise the tracking and
verification lazy budgets to 12,600 and 7,300 bytes.

Benchmarks showed the cost reason never paid: the cheaper worker needed far
more turns and left more defects, so verification and corrections cost more
than the writing saved. A reported price ratio no longer fires the Writer
rule and the runtime no longer injects it into the prompt. Configured
per-agent models still apply to writers launched for other reasons, and the
price-ratio library stays for later experiments.

A verifier created its scratch copy and Node's compile cache inside the
workspace, leaving untracked directories the orchestrator then had to ask
to delete. Create the scratch with mktemp -d under the system temp
directory, keep NODE_COMPILE_CACHE there, and leave no new file in the
workspace.

In pi -p and pi --mode json the runtime already rejects background
subagents, but the prompt still announced the configured policy, so the
orchestrator tried background launches and took the parallel-writer path.
Single-shot host modes now render the policy as off. Also ask inline and
delegated work to update the help text and docs for any changed command,
option, or message.

Only the delegated worker had a TRIANGULATE step; inline work went from
GREEN straight to refactor, which left tests thinner than Codex's. Both
paths now add at least two edge cases beyond the request's examples after
GREEN (a boundary or empty value, and an invalid input followed by a valid
one) and run them once, without a RED run since they are expected to pass.

A fixed count of generic edge cases padded the tests without catching the
defects the blind reviews found, which all came from the request's rules or
from existing commands the change touched. Write one RED test per requested
rule covering the cases that rule names, and one test per touched existing
command or option proving its previous behavior still holds; nothing else.

…ason

With background subagents unavailable, the parallelism reason still fired
and the writers ran one after another in task mode, adding subagent overhead
without any parallel work. The core Writer rule now requires the units to
launch together in background; otherwise the parent works inline.

… shared

Every blind-reviewed solution changed how a sibling subcommand rejected a
newly shared option without a test. Name the concrete case in the PRESERVE
rule: test that sibling subcommands that did not accept the option still
reject it with the same error.

… in risk

The same task was classified high risk in one run and not in another, and
the cited reason was the change the user asked for, while the regression
verify found was in shared code nobody asked to change. High-risk item 3
now says requested changes are not high risk by themselves and unrequested
breaks in shared code are. The parser-specific PRESERVE case did not work
and is replaced by a general definition of touched behavior.

The nan provider is registered by a gentle-pi extension that subagent
children never loaded, so any subagent routed to a nan model failed at
start with model not found. Children now receive the nan provider
extension next to the context and safety extensions.

Runs that changed options shared with other existing commands declared
Risk: none, skipped verification, and one shipped a silent regression in
a sibling command. Before writing Risk: none, the orchestrator now checks
whether its diff changes code that unmentioned existing behavior also
uses, which counts as high-risk item 3.

…riter rule

Address review: the readme reference and the gentle-ai skill still described
size-based writer delegation, and the feature document committed local
absolute paths.