@azure-id/orc 1.1.0 → 1.2.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +2663 -2437
- package/README.md +665 -694
- package/bin/cli.js +658 -2
- package/bin/verify-contracts.js +63 -1
- package/bin/verify-package.js +573 -568
- package/package.json +1 -1
- package/templates/hooks/README.md +202 -0
- package/templates/hooks/orc-statusline.js +552 -9
- package/templates/skills/_shared/phases/execution.md +2 -0
- package/templates/skills/_shared/return-validation.md +222 -145
- package/templates/skills/orc/SKILL.md +241 -237
- package/templates/skills/orc-doc/SKILL.md +484 -480
- package/templates/skills/orc-fast/SKILL.md +216 -214
- package/templates/skills/orc-mini/SKILL.md +246 -243
- package/templates/skills/orc-quick/SKILL.md +347 -345
- package/templates/skills/orc-wiki/SKILL.md +174 -170
|
@@ -1,145 +1,222 @@
|
|
|
1
|
-
# Return validation (every lane, every subagent return)
|
|
2
|
-
|
|
3
|
-
Canonical procedure for validating a spawned agent's return. Every ORC lane
|
|
4
|
-
(full, mini, fast, wiki, diy) runs this on EVERY return; a malformed return is
|
|
5
|
-
a failure (requeue/re-dispatch with reason — lane sets the retry cap).
|
|
6
|
-
|
|
7
|
-
##
|
|
8
|
-
|
|
9
|
-
|
|
10
|
-
|
|
11
|
-
|
|
12
|
-
|
|
13
|
-
|
|
14
|
-
|
|
15
|
-
|
|
16
|
-
|
|
17
|
-
|
|
18
|
-
-
|
|
19
|
-
|
|
20
|
-
|
|
21
|
-
|
|
22
|
-
|
|
23
|
-
|
|
24
|
-
|
|
25
|
-
|
|
26
|
-
|
|
27
|
-
|
|
28
|
-
|
|
29
|
-
|
|
30
|
-
|
|
31
|
-
|
|
32
|
-
|
|
33
|
-
|
|
34
|
-
|
|
35
|
-
|
|
36
|
-
|
|
37
|
-
|
|
38
|
-
|
|
39
|
-
|
|
40
|
-
|
|
41
|
-
|
|
42
|
-
|
|
43
|
-
|
|
44
|
-
|
|
45
|
-
|
|
46
|
-
|
|
47
|
-
|
|
48
|
-
|
|
49
|
-
|
|
50
|
-
|
|
51
|
-
|
|
52
|
-
|
|
53
|
-
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
|
|
60
|
-
|
|
61
|
-
|
|
62
|
-
|
|
63
|
-
|
|
64
|
-
|
|
65
|
-
|
|
66
|
-
|
|
67
|
-
|
|
68
|
-
|
|
69
|
-
|
|
70
|
-
|
|
71
|
-
|
|
72
|
-
|
|
73
|
-
|
|
74
|
-
|
|
75
|
-
|
|
76
|
-
|
|
77
|
-
|
|
78
|
-
|
|
79
|
-
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
|
|
100
|
-
|
|
101
|
-
|
|
102
|
-
|
|
103
|
-
|
|
104
|
-
|
|
105
|
-
|
|
106
|
-
|
|
107
|
-
|
|
108
|
-
|
|
109
|
-
|
|
110
|
-
|
|
111
|
-
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
`
|
|
115
|
-
|
|
116
|
-
|
|
117
|
-
|
|
118
|
-
|
|
119
|
-
|
|
120
|
-
|
|
121
|
-
|
|
122
|
-
|
|
123
|
-
|
|
124
|
-
|
|
125
|
-
|
|
126
|
-
|
|
127
|
-
|
|
128
|
-
|
|
129
|
-
|
|
130
|
-
|
|
131
|
-
(`
|
|
132
|
-
|
|
133
|
-
|
|
134
|
-
|
|
135
|
-
|
|
136
|
-
|
|
137
|
-
|
|
138
|
-
|
|
139
|
-
|
|
140
|
-
|
|
141
|
-
|
|
142
|
-
return
|
|
143
|
-
|
|
144
|
-
|
|
145
|
-
|
|
1
|
+
# Return validation (every lane, every subagent return)
|
|
2
|
+
|
|
3
|
+
Canonical procedure for validating a spawned agent's return. Every ORC lane
|
|
4
|
+
(full, mini, fast, wiki, diy) runs this on EVERY return; a malformed return is
|
|
5
|
+
a failure (requeue/re-dispatch with reason — lane sets the retry cap).
|
|
6
|
+
|
|
7
|
+
## 0. Is the previous attempt still ALIVE? (v1.2.0) — BEFORE anything else
|
|
8
|
+
|
|
9
|
+
> **`a lane that re-dispatches over a live attempt` has broken this contract.**
|
|
10
|
+
|
|
11
|
+
**A Task error does not kill the agent behind it.** Claude Code's tool call can
|
|
12
|
+
fail, time out, or be cut off mid-turn while the subagent it started keeps
|
|
13
|
+
running — and keeps writing files. Every rule below this line ends in
|
|
14
|
+
"re-dispatch", and every one of them silently assumed a failed call meant a dead
|
|
15
|
+
agent. It does not.
|
|
16
|
+
|
|
17
|
+
What that costs, measured: one graded `/orc-quick` entry put THREE
|
|
18
|
+
`orc-executor-opus-5-low` agents on the SAME task — 50m19s, 115m22s and
|
|
19
|
+
100m53s, **266 minutes of Opus 5 for one authorised dispatch**, all editing the
|
|
20
|
+
same files, inside a 2h04m window. The second was dispatched 4m19s after the
|
|
21
|
+
first, while the first was still working. The hook had recorded all three; no
|
|
22
|
+
lane had ever read that record.
|
|
23
|
+
|
|
24
|
+
### The rule
|
|
25
|
+
|
|
26
|
+
Before ANY re-dispatch, requeue or repair round — and before the first dispatch
|
|
27
|
+
of a resumed run — run:
|
|
28
|
+
|
|
29
|
+
```
|
|
30
|
+
orc run inflight --json # 0 clear · 1 in-flight · 2 unknown
|
|
31
|
+
```
|
|
32
|
+
|
|
33
|
+
| exit | meaning | what the lane does |
|
|
34
|
+
|---|---|---|
|
|
35
|
+
| 0 | provably nothing in flight | dispatch |
|
|
36
|
+
| 1 | ≥1 dispatch has not returned | **REFUSE. Name the agent, the task and its age.** Ask the user. |
|
|
37
|
+
| 2 | cannot prove either way | **REFUSE by default.** Say why, and let the USER decide. |
|
|
38
|
+
|
|
39
|
+
**Exit 2 refuses, and that is deliberate.** Everywhere else in ORC an absent
|
|
40
|
+
reading is treated as absent and never blocks — `orc usage check` exit 2 never
|
|
41
|
+
stops a run, an UNCHECKABLE pact never raises the exit code. This is the one
|
|
42
|
+
place the default inverts, because the two outcomes are not symmetrical: a
|
|
43
|
+
wrongly-refused dispatch costs one question, and a wrongly-issued one costs a
|
|
44
|
+
second Opus agent for an hour. Refusing on `unknown` is the cheap error.
|
|
45
|
+
|
|
46
|
+
### An interrupted turn is UNKNOWN, never FAILED
|
|
47
|
+
|
|
48
|
+
A usage limit, an API error, a dropped connection or a `Ctrl+C` between a
|
|
49
|
+
dispatch and its return says **nothing** about the agent. Treat it as §0 exit 2
|
|
50
|
+
and ask. A lane that classifies an interruption as a failure re-dispatches into
|
|
51
|
+
a live agent, gets interrupted again sooner because it is now paying twice, and
|
|
52
|
+
the loop tightens on itself — which is exactly how the 266-minute entry
|
|
53
|
+
happened.
|
|
54
|
+
|
|
55
|
+
### What refusing looks like
|
|
56
|
+
|
|
57
|
+
```
|
|
58
|
+
⛔ 1 dispatch is still in flight — not re-dispatching.
|
|
59
|
+
|
|
60
|
+
orc-executor-opus-5-low started 4m ago
|
|
61
|
+
"Fix approval flow defects"
|
|
62
|
+
|
|
63
|
+
A Task error does not kill the agent behind it. It may still be writing.
|
|
64
|
+
1. wait for it 2. dispatch anyway (2 agents on one task) 3. stop
|
|
65
|
+
```
|
|
66
|
+
|
|
67
|
+
Option 2 must always be offered and never be the default: the user is allowed
|
|
68
|
+
to overrule this, and an unreadable sidecar must never trap a run.
|
|
69
|
+
|
|
70
|
+
### The evidence, and its one honest limit
|
|
71
|
+
|
|
72
|
+
`orc run inflight` reads the pending sidecar that `orc-trace.js` writes on every
|
|
73
|
+
`SPAWN`, cross-checked against the trace's own SPAWN/RETURN balance. It reports
|
|
74
|
+
`unknown` — never `clear` — when the sidecar is missing, unreadable, or holds
|
|
75
|
+
only records older than six hours, and when the sidecar and the trace disagree.
|
|
76
|
+
**Unknown is not zero.**
|
|
77
|
+
|
|
78
|
+
It cannot see an **ad-hoc** dispatch (`/orc-quick` recon, model+effort rather
|
|
79
|
+
than a pinned `orc-*` agent): the hook writes no `SPAWN` for one, so no record
|
|
80
|
+
exists. Those are read-only and short, so the exposure is small — but the limit
|
|
81
|
+
is stated rather than papered over, and a lane must not report `clear` as proof
|
|
82
|
+
that an ad-hoc read is finished.
|
|
83
|
+
|
|
84
|
+
## 1. Contract shape
|
|
85
|
+
|
|
86
|
+
The return must carry every field its agent contract names. Missing or extra
|
|
87
|
+
shape = malformed. Never repair a return yourself; re-dispatch.
|
|
88
|
+
|
|
89
|
+
## 2. Claimed-vs-actual model (tier-downgrade check)
|
|
90
|
+
|
|
91
|
+
Every return carries:
|
|
92
|
+
|
|
93
|
+
- `actual_model` — quoted VERBATIM from the agent's system-prompt model-id
|
|
94
|
+
line, never inferred (`unknown` when no such line exists)
|
|
95
|
+
- `actual_effort` — the agent's `$CLAUDE_EFFORT` value
|
|
96
|
+
|
|
97
|
+
Compare both against what the dispatch expected (the agent NAME encodes it).
|
|
98
|
+
Append the `VERIFY` trace line with the comparison; any mismatch is surfaced
|
|
99
|
+
to the user as a ⛔ DOWNGRADE — never silently accepted. (A subagent can't
|
|
100
|
+
exceed the MAIN session's tier, so a downgrade usually means the main session
|
|
101
|
+
is on the wrong model.)
|
|
102
|
+
|
|
103
|
+
## 2b. A FOREIGN return — the SUBSTITUTION check (v0.50.0)
|
|
104
|
+
|
|
105
|
+
A foreign worker (`orc extra dispatch`, `_shared/extra-dispatch.md`) is not a
|
|
106
|
+
Claude subagent. It has no injected system-prompt model-id line, so **it cannot
|
|
107
|
+
carry `actual_model`** — and §2 must not be faked for it. A foreign return that
|
|
108
|
+
claimed an `actual_model` would be claiming evidence that does not exist.
|
|
109
|
+
|
|
110
|
+
It carries instead, and every one of these is quoted from the wire rather than
|
|
111
|
+
assumed:
|
|
112
|
+
|
|
113
|
+
- `engine` (`api` | `claude-shim` | `cli`), `provider`, `profile`
|
|
114
|
+
- `model_requested` — what the route row asked for
|
|
115
|
+
- **`model_reported`** — the `model` field the endpoint echoed back
|
|
116
|
+
- `usage` — the four token kinds, never blended, **or `null`**
|
|
117
|
+
|
|
118
|
+
**`model_reported != model_requested` is ⛔ SUBSTITUTION**, surfaced to the user
|
|
119
|
+
exactly as ⛔ DOWNGRADE is today and never silently accepted. It is the only
|
|
120
|
+
defence against an aggregator quietly serving something else. `unknown` is a
|
|
121
|
+
valid, honest value and is reported as `unknown` — **never as a match**.
|
|
122
|
+
|
|
123
|
+
**A clean model check is not a clean answer.** An aggregator's *provider-level*
|
|
124
|
+
fallback is on by default and it PRESERVES the model id, so the substitution
|
|
125
|
+
check reads clean while the code went to a different company. Engine `api`
|
|
126
|
+
records the response's `provider` echo and reports **⚠ REROUTE**; the other two
|
|
127
|
+
engines cannot see it at all, and their `served_by_note` says so. An absent
|
|
128
|
+
measurement is never a pass.
|
|
129
|
+
|
|
130
|
+
**`usage: null` is not four zeros.** A worker that reported no token counts
|
|
131
|
+
(engine `cli` frequently) returns `null` plus a note. `{0,0,0,0}` would tell
|
|
132
|
+
`/orc-budget` the run was free. Engine `api`'s `cache_write: 0` is the opposite
|
|
133
|
+
case — a real measurement — so the two must never be normalised together.
|
|
134
|
+
|
|
135
|
+
**The fence is per-engine, and the return says which one it had.** Engine `api`
|
|
136
|
+
ENFORCES `declared_files`; engines `claude-shim` and `cli` ASK. A return
|
|
137
|
+
carrying `fence: {declared_files: false}` means the list was an instruction, not
|
|
138
|
+
a rule — treat §6 below as the only real check, and say so to the user rather
|
|
139
|
+
than reporting a constraint that was never applied.
|
|
140
|
+
|
|
141
|
+
**A RESUMED foreign dispatch owes three more fields** (v0.54.0). The dispatch
|
|
142
|
+
return sets `resume_expected: true`, so the obligation is never inferred:
|
|
143
|
+
|
|
144
|
+
- **`resume_state`** — `continued` · `restarted` · `no-op`. Absent on a slice
|
|
145
|
+
with no `resumed_from` is correct; **absent on a resume slice is MALFORMED.**
|
|
146
|
+
A return claiming `restarted` while `preexisting[]` was non-empty is a
|
|
147
|
+
FINDING, not a failure — it is how `/orc-retro` learns which providers ignore
|
|
148
|
+
a resume preamble, so surface it rather than treating it as a bad return.
|
|
149
|
+
- **`preexisting_read[]`** — which pre-existing files the worker actually
|
|
150
|
+
opened. Quoted like `wiki_used`: **what it did, never what the dispatcher
|
|
151
|
+
assumed.** An EMPTY list on a resume whose `preexisting[]` was not empty is an
|
|
152
|
+
honest and informative return — it says the worker ignored the preamble — and
|
|
153
|
+
it must be surfaced, never dropped.
|
|
154
|
+
- **`journal_fidelity`** — relayed from the dispatch return (`per-turn` |
|
|
155
|
+
`streamed-opaque`), so a validator never reports `streamed-opaque` evidence as
|
|
156
|
+
if it had per-turn tool attribution.
|
|
157
|
+
|
|
158
|
+
Everything else in this file applies to a foreign return unchanged: the
|
|
159
|
+
honest-status rules, the pattern attestation, the TDD attestation, the wiki
|
|
160
|
+
attestation, and above all **§6, the worktree delta** — which is engine-blind
|
|
161
|
+
because it reads the worktree rather than the return, and is therefore what
|
|
162
|
+
makes a foreign executor safe at all.
|
|
163
|
+
|
|
164
|
+
## 3. Honest-status rules (executor returns)
|
|
165
|
+
|
|
166
|
+
- `status=done` on a stack with a runnable build/test REQUIRES `evidence`
|
|
167
|
+
{command, exit_code, tail} quoted VERBATIM; a missing block or a false
|
|
168
|
+
`no_runner_detected` is malformed.
|
|
169
|
+
- `done` with a non-empty `unmet[]` is `partial` — treat it as such.
|
|
170
|
+
|
|
171
|
+
## 4. Pattern attestation (when a `pattern` was injected)
|
|
172
|
+
|
|
173
|
+
A task that received a `pattern` slice must return `invariants_checked: true`
|
|
174
|
+
plus the matching `pattern_version`; false/absent on a pattern task is
|
|
175
|
+
malformed.
|
|
176
|
+
|
|
177
|
+
## 5. TDD attestation (when a `tdd_spec` was injected — v0.33.0)
|
|
178
|
+
|
|
179
|
+
A task whose slice carried a `tdd_spec` must return `tdd_state: green|red` —
|
|
180
|
+
`green` only with the passing run quoted in `evidence`; `status=done` with
|
|
181
|
+
`tdd_state: red` (or an absent field) is malformed. `red` is an HONEST return:
|
|
182
|
+
the lane runs its repair loop up to `tdd_loop_max`, then STOPS with the red
|
|
183
|
+
report — never re-dispatch past the cap.
|
|
184
|
+
|
|
185
|
+
## 5b. Wiki attestation (when wiki content or page pointers were injected — v0.41.0)
|
|
186
|
+
|
|
187
|
+
A task whose slice carried wiki material must return **`wiki_used`** — the doc
|
|
188
|
+
paths it ACTUALLY read, or `none`. Quoted like `actual_model`: what the agent
|
|
189
|
+
did, never what the dispatcher assumed.
|
|
190
|
+
|
|
191
|
+
`none` is a valid and INFORMATIVE return, not a failure: it says the pages were
|
|
192
|
+
not useful or were ignored. Record it and surface it — a wiki whose pages are
|
|
193
|
+
shipped into every slice and read by nobody is the failure mode this field
|
|
194
|
+
exists to make visible, and it is invisible if `none` is quietly dropped. Absent
|
|
195
|
+
on a slice that carried wiki material is malformed. Not required otherwise.
|
|
196
|
+
|
|
197
|
+
## 6. Worktree delta (post-wave, every lane that dispatches executors)
|
|
198
|
+
|
|
199
|
+
Compare `git status --short` before and after each dispatch. A path that
|
|
200
|
+
appears, disappears, or **reverts** and is absent from that task's
|
|
201
|
+
`declared_files` is a slice violation regardless of what the return said —
|
|
202
|
+
including a file that became LESS modified, which is how a destructive `git`
|
|
203
|
+
command inside a slice disguises itself as a clean tree. `actual_files` is a
|
|
204
|
+
CLAIM; the worktree is the EVIDENCE. An unexplained delta gates the wave: name
|
|
205
|
+
it, attribute it, and get a decision before closing.
|
|
206
|
+
|
|
207
|
+
**On a RESUMED task the "before" side of the delta is the JOURNAL BASELINE**
|
|
208
|
+
(`orc extra reconcile`, `_shared/extra-dispatch.md`), not the state at the top of
|
|
209
|
+
this wave. A file the previous attempt created is already in the tree and is
|
|
210
|
+
**not** an unexplained delta — it is explained, by the journal, by name. Without
|
|
211
|
+
that the first resumed wave trips its own gate on the work it just recovered.
|
|
212
|
+
|
|
213
|
+
## 7. Gotcha capture (repair loops only — v0.40.0)
|
|
214
|
+
|
|
215
|
+
A return that closes a repair loop (`tdd_state` went red → green, a drift
|
|
216
|
+
round resolved, a reviewer P0/P1 was fixed in-run) carries
|
|
217
|
+
`gotcha_recorded` — either the entry body (`trigger`, `symptom`, `cause`,
|
|
218
|
+
`fix`, `scope`) or `none` with a one-line reason. Absent on a repair-closing
|
|
219
|
+
return is malformed. It is NOT required on a return that never repaired
|
|
220
|
+
anything, and a loop that hit its cap and STOPPED must return `none` — an
|
|
221
|
+
unsolved failure is not a gotcha. The agent RETURNS the body; the
|
|
222
|
+
orchestrator writes the file. See `gotchas.md`.
|