feat(aidd-orchestrator): add an interactive SDLC mode - #967

…ate becomes reachable

Four defects the case battery surfaced in zones the interactive mode did
not own. Each left the mode promising something the flow could not keep.

Frame's spec and Check's review had no committer, so a request could open
without the contract it was built from or the review that cleared it. Each
zone now commits what it wrote before handing over.

Whether an end to end journey was required rested on a file no zone
reference points at, so a run that obeyed the reading rule skipped it. The
project's testing memory decides, and nothing else does.

`planning-ready` decided whether a spec is written, and therefore how often
a person is consulted, with no test behind it. Three tests now hold it: the
source names the artifact to change, states the observable outcome, and
leaves open no decision that changes what is built.

The request gate read `no actionable finding`, and a strict checker always
had one, so two earlier runs looped four times each and never reached it.
Once the review has cleared, only a finding that contradicts the contract
reopens a zone; any other is carried onto the request as a note. A pass
budget was rejected: the number would be arbitrary and indefensible.

Eleven headless runs, one per claim and one per branch of each claim. The
direct branch held on the same request that twenty earlier runs had settled
there, so the third test is not stricter than the flow it describes.

Co-Authored-By: Claude Opus 5 <[email protected]>
AIDD-Session-Id: bf713ca4-3c1b-491d-ac0a-4d4533330b5c

The mode sentence names one condition instead of four, the stop rule moves
next to the zone reading it, and the diagram carries each zone's pause so the
flow and the prose cannot drift apart.

The onboard tier note no longer names a skill from another plugin; it states
the rule any autonomous skill follows.

Co-Authored-By: Claude Opus 5 <[email protected]>
AIDD-Session-Id: bf713ca4-3c1b-491d-ac0a-4d4533330b5c

…not judged

The planning-ready test becomes its own sentence, so a zone can answer it
instead of inferring it from a clause. The journey runs when the plan calls
for one and never otherwise, which moves the decision to the artifact that
can state it.

Co-Authored-By: Claude Opus 5 <[email protected]>
AIDD-Session-Id: bf713ca4-3c1b-491d-ac0a-4d4533330b5c

The two sentences governing a recurring finding both claimed it: a cosmetic
finding a pass had already acted on was a non-blocking note under one and a
blocking stop under the other. Three independent readers answered that the
text did not decide.

The reopening rule now runs first, and the recurrence rule speaks of a finding
that would be routed a second time, so a note never reaches it. Four readers
re-traced both paths on the corrected text and none answered undecidable.

Co-Authored-By: Claude Opus 5 <[email protected]>
AIDD-Session-Id: bf713ca4-3c1b-491d-ac0a-4d4533330b5c

A Codex run of the interactive flow answered two of its three gates with
"approved, with corrections". The text had no reading for that: a correction is a
finding against the artifact, so the letter of it sent a one-line fix back
through the zone and presented the artifact again. The run advanced instead,
which is the useful behaviour and not what the text said.

Three readers traced the three answers a pause can get — approved, approved with
corrections, refused — and none answered undecidable. The run itself, including
the three gates exercised with a person at them, is recorded beside the batch.

Co-Authored-By: Claude Opus 5 <[email protected]>
AIDD-Session-Id: bf713ca4-3c1b-491d-ac0a-4d4533330b5c

The interactive gates had never been executed, only read: a headless run cannot
answer them without inventing the approval the gate exists to collect. Two Codex
runs close that, one per mode, on the same issue from the same commit.

Both reached the same fix. The gates added no quality on it, and the person at
the first one made it worse, so the comparison is recorded as measured rather
than as the result the design expected.

Co-Authored-By: Claude Opus 5 <[email protected]>
AIDD-Session-Id: bf713ca4-3c1b-491d-ac0a-4d4533330b5c

#910 asks that documentation explain both modes and their default, and the
review had stripped every mention. The description, the plugin table and the
README row say autonomous or supervised, and the catalogs follow from the
description.

The word a request carries stays `interactive`, which the argument hint already
names. The README row names it too, so the word a reader meets is the word they
can type.

Co-Authored-By: Claude Opus 5 <[email protected]>
AIDD-Session-Id: bf713ca4-3c1b-491d-ac0a-4d4533330b5c

`## Behavior` had grown from 84 words to 313, almost all of it the mode: nine
sentences, each one a run that failed without it. Compressing them would have
traded a proven case for a shorter page — an earlier pass already lost the plural
in "every wait a step takes" that way and let brainstorm skip its own pause.

So they move whole into `references/mode.md`, a section per mode, and the router
keeps the dispatch alone at 77 words. One redundancy went with the structure:
inside a section named Interactive, "dispatch in interactive" says nothing.

Three readers starting from SKILL.md alone found the reference and applied it:
the five mode cases, the three pauses in order, the zero pauses of auto with its
three exceptions, and where a refused outcome routes.

Co-Authored-By: Claude Opus 5 <[email protected]>
AIDD-Session-Id: bf713ca4-3c1b-491d-ac0a-4d4533330b5c

Four paragraphs of run-earned sentences read as prose when every one of them is a
constraint to check. They become bulleted rules under Behavior, Auto, Interactive
and Refusal, so a reader meets one rule at a time.

Two rules went in the process. "Dispatch only the steps … and run the rest as
`## Auto` says" pointed at a section of its own file, which reads as nothing: it
now says that everything else runs without pausing. And the opening pair, naming
the three pauses then saying to present and wait, was one rule written twice.

Five readers confirm the format did not turn rules into a procedure: the five mode
cases, the three pauses in order, no bullet contradicting another, an approval
carrying a correction, and — step by step — that clarifying and formalizing are
the contract pause rather than two more, while implementing and repairing pause
for nothing.

Co-Authored-By: Claude Opus 5 <[email protected]>
AIDD-Session-Id: bf713ca4-3c1b-491d-ac0a-4d4533330b5c

Eight sentences in one paragraph, three of them added by this branch, and two
routings living in two files with nothing saying they were different things: a
checker finding routes by what it is about, a refusal the user gives at a pause
routes by artifact. A reader had to infer that from the word "finding" meaning
two things.

The zone now reads as rules under Behavior, Findings and Handoff, and Findings
opens by saying which of the two it governs and where the other one lives.

Three readers crossed the two on four cases — an implementation finding before
the review clears, a contract refusal at the outcome pause, a contract
contradiction after it cleared, and a plan refusal. All four landed in the same
zone for all three readers, each through the right file, and none found the two
routings in conflict.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01VE16AxC1V9wAMGmGgefe4z
AIDD-Session-Id: bf713ca4-3c1b-491d-ac0a-4d4533330b5c

…ring it

Every earlier run was headless, so the three gates were only read: a headless run
cannot answer a pause without inventing the approval the pause exists to collect.
This one stopped at each pause, reported what it presented, and continued from a
real answer — approve, approve with a correction, refuse, refuse again.

The source was deliberately not planning-ready, so the clarify then formalize
path ran. That path had never been counted: every headless interactive run used a
planning-ready source and went through Frame directly, which is why the pause
counting defect survived 38 of them and surfaced only when asked about.

Five pauses on three artifacts, two step waits correctly held not to be pauses,
ten commits, seventeen tests green, no pull request. The run also volunteered two
faults of its own conduct — probe files outside the sandbox, and a mutation
measured against a reverted tree — neither of which the reviewer had caught.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01VE16AxC1V9wAMGmGgefe4z
AIDD-Session-Id: bf713ca4-3c1b-491d-ac0a-4d4533330b5c