loki-mode 8.0.3 → 8.2.0
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/README.md +54 -10
- package/SKILL.md +2 -2
- package/VERSION +1 -1
- package/autonomy/completion-council.sh +45 -4
- package/autonomy/council-v2.sh +114 -16
- package/autonomy/grill.sh +32 -0
- package/autonomy/lib/done-recognition.sh +16 -4
- package/autonomy/lib/fast_verify.py +346 -0
- package/autonomy/lib/no_mock_scan.py +21 -1
- package/autonomy/lib/prd-enrich.sh +12 -1
- package/autonomy/lib/proof-generator.py +173 -6
- package/autonomy/lib/proof-template.html +77 -12
- package/autonomy/lib/proof-verify.py +52 -0
- package/autonomy/loki +246 -7
- package/autonomy/run.sh +14 -0
- package/autonomy/verify.sh +12 -6
- package/dashboard/__init__.py +1 -1
- package/events/emit.sh +105 -18
- package/loki-ts/dist/loki.js +319 -316
- package/mcp/__init__.py +30 -12
- package/mcp/server.py +150 -0
- package/package.json +1 -1
- package/plugins/loki-mode/.claude-plugin/plugin.json +1 -1
- package/providers/claude.sh +72 -0
- package/providers/codex.sh +13 -0
- package/providers/loader.sh +10 -4
- package/providers/model_catalog.json +114 -17
- package/providers/models.sh +26 -3
- package/providers/opencode.sh +145 -0
- package/references/design-archetypes.md +85 -0
package/README.md
CHANGED
|
@@ -15,7 +15,7 @@ _The free, source-available autonomous coding agent by [Autonomi](https://www.au
|
|
|
15
15
|
|
|
16
16
|
[Website](https://www.autonomi.dev/) | [Documentation](wiki/Home.md) | [Installation](docs/INSTALLATION.md) | [Changelog](CHANGELOG.md) | [Purple Lab -- deprecated v7.44.0](#purple-lab)
|
|
17
17
|
|
|
18
|
-
**Current release: v8.
|
|
18
|
+
**Current release: v8.2.0**
|
|
19
19
|
|
|
20
20
|
</div>
|
|
21
21
|
|
|
@@ -25,6 +25,37 @@ _The free, source-available autonomous coding agent by [Autonomi](https://www.au
|
|
|
25
25
|
|
|
26
26
|
---
|
|
27
27
|
|
|
28
|
+
## Already have a codebase? Start read-only.
|
|
29
|
+
|
|
30
|
+
Most agents are built to create new apps. The harder, more valuable problem is
|
|
31
|
+
the ten-year-old repo that pays the bills. Loki works on both, and on an
|
|
32
|
+
existing codebase it starts by **changing nothing**:
|
|
33
|
+
|
|
34
|
+
```bash
|
|
35
|
+
loki modernize heal ./your-repo --assess # read-only. no writes, no commits.
|
|
36
|
+
loki modernize heal ./your-repo --assess --json # same, machine-readable
|
|
37
|
+
```
|
|
38
|
+
|
|
39
|
+
You get a modernization readiness report: language mix, a 4-level maturity
|
|
40
|
+
rating, technical-debt signals (test coverage, TODO density, oversized files,
|
|
41
|
+
dependency staleness), and a **ranked list of where to start** -- ordered by
|
|
42
|
+
blast radius, so the first change is the one least likely to break something.
|
|
43
|
+
|
|
44
|
+
Then, if you want it to act:
|
|
45
|
+
|
|
46
|
+
```bash
|
|
47
|
+
loki modernize heal ./your-repo --strict # block ALL behavioral change without approval
|
|
48
|
+
loki modernize heal ./your-repo --phase archaeology # extract knowledge only
|
|
49
|
+
loki modernize heal ./your-repo --compliance healthcare # or fintech | government
|
|
50
|
+
```
|
|
51
|
+
|
|
52
|
+
The healing pipeline runs in phases -- archaeology, stabilize, isolate,
|
|
53
|
+
modernize, validate -- and the validate phase checks **behavioral equivalence
|
|
54
|
+
against the pre-change baseline**, not just that the tests are green. Friction
|
|
55
|
+
points (the weird code that exists for a reason nobody remembers) are cataloged
|
|
56
|
+
before anything touches them, because in a legacy system the strange code is
|
|
57
|
+
usually load-bearing.
|
|
58
|
+
|
|
28
59
|
## The Evidence Receipt: don't trust the agent, check it
|
|
29
60
|
|
|
30
61
|
Every coding agent tells you it finished. Loki hands you something you can
|
|
@@ -239,18 +270,31 @@ This needs Bun on your PATH (the SDK loop runs on the Bun runtime). `loki doctor
|
|
|
239
270
|
|
|
240
271
|
**With a coding-agent CLI.** The classic path, and still the default: Loki drives a separate CLI (Claude Code is the recommended one) plus a couple of common tools on your PATH.
|
|
241
272
|
|
|
242
|
-
**With a different model or provider.** Loki is not tied to Anthropic
|
|
243
|
-
|
|
244
|
-
|
|
245
|
-
|
|
273
|
+
**With a different model or provider.** Loki is not tied to Anthropic, but
|
|
274
|
+
*how* you reach another model depends on which API the endpoint speaks. There
|
|
275
|
+
are two routes, and picking the wrong one fails confusingly.
|
|
276
|
+
|
|
277
|
+
*Route 1 -- OpenAI-shaped endpoints (OpenRouter, and most hosted open models).*
|
|
278
|
+
Use a provider that speaks that API natively. `aider` and `cline` both do, and
|
|
279
|
+
Loki now defaults them to open-weight models rather than Claude:
|
|
246
280
|
|
|
247
281
|
```bash
|
|
248
|
-
|
|
249
|
-
export
|
|
250
|
-
|
|
251
|
-
export LOKI_MODEL_OVERRIDE=<model-id-from-openrouter.ai/models>
|
|
252
|
-
loki start prd.md
|
|
282
|
+
loki provider set aider
|
|
283
|
+
export OPENROUTER_API_KEY=sk-or-...
|
|
284
|
+
loki start prd.md # defaults to deepseek-v3.2
|
|
253
285
|
|
|
286
|
+
export LOKI_AIDER_MODEL=openrouter/z-ai/glm-4.6 # or pick your own
|
|
287
|
+
```
|
|
288
|
+
|
|
289
|
+
OpenRouter serves **only** the OpenAI-shaped `/v1/chat/completions`; it has no
|
|
290
|
+
Anthropic `/v1/messages` endpoint. Pointing `ANTHROPIC_BASE_URL` at it does not
|
|
291
|
+
work, which earlier versions of this README incorrectly suggested.
|
|
292
|
+
|
|
293
|
+
*Route 2 -- Anthropic-protocol gateways.* `ANTHROPIC_BASE_URL` routes Claude
|
|
294
|
+
Code itself, so the endpoint must speak the Anthropic Messages API. LiteLLM,
|
|
295
|
+
Bedrock proxies, and self-hosted gateways can:
|
|
296
|
+
|
|
297
|
+
```bash
|
|
254
298
|
# Ollama, fully local (no API key, no per-token cost)
|
|
255
299
|
export ANTHROPIC_BASE_URL=http://localhost:11434/v1
|
|
256
300
|
export LOKI_MODEL_OVERRIDE=<model you have pulled, e.g. the output of `ollama list`>
|
package/SKILL.md
CHANGED
|
@@ -3,7 +3,7 @@ name: loki-mode
|
|
|
3
3
|
description: Autonomous spec-driven build system with a built-in trust layer. It does not call work done until it is verified (RARV-C closure loop, 8 quality gates, completion council, verified-completion evidence gate). Triggers on "Loki Mode". Takes a spec (PRD, GitHub issue, OpenAPI doc, etc.) to deployed product with minimal human intervention. Provider-agnostic. Requires --dangerously-skip-permissions flag.
|
|
4
4
|
---
|
|
5
5
|
|
|
6
|
-
# Loki Mode v8.0
|
|
6
|
+
# Loki Mode v8.2.0
|
|
7
7
|
|
|
8
8
|
**You are an autonomous agent. You make decisions. You do not ask questions. You do not stop.**
|
|
9
9
|
|
|
@@ -469,4 +469,4 @@ See `CHANGELOG.md` entries [7.5.7], [7.5.8], [7.5.13] for the per-fix list and r
|
|
|
469
469
|
|
|
470
470
|
---
|
|
471
471
|
|
|
472
|
-
**v8.0
|
|
472
|
+
**v8.2.0 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**
|
package/VERSION
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
8.0
|
|
1
|
+
8.2.0
|
|
@@ -2267,11 +2267,17 @@ PYEOF
|
|
|
2267
2267
|
if [ -z "$_nm_cc_dir" ] || [ ! -f "$_nm_scanner" ]; then
|
|
2268
2268
|
_nm_status="INCONCLUSIVE:scanner_unavailable"
|
|
2269
2269
|
else
|
|
2270
|
+
# STDIN, not an env var: a single env string is capped at
|
|
2271
|
+
# MAX_ARG_STRLEN (131071 bytes) on Linux, so a large changed-file
|
|
2272
|
+
# union makes execve fail with E2BIG and this gate silently
|
|
2273
|
+
# degrades to inconclusive (pass-through). macOS has no per-string
|
|
2274
|
+
# cap, so that failure mode was Linux-only.
|
|
2270
2275
|
_nm_status=$(
|
|
2271
|
-
|
|
2272
|
-
|
|
2273
|
-
|
|
2274
|
-
|
|
2276
|
+
printf '%s\n' "$_nm_files" | {
|
|
2277
|
+
_NM_TREE="." \
|
|
2278
|
+
python3 -I "$_nm_scanner" 2>/dev/null \
|
|
2279
|
+
|| echo "INCONCLUSIVE:detector_error"
|
|
2280
|
+
}
|
|
2275
2281
|
)
|
|
2276
2282
|
fi
|
|
2277
2283
|
case "$_nm_status" in
|
|
@@ -3135,6 +3141,25 @@ ISSUES: CRITICAL:description (optional, one per line per issue)"
|
|
|
3135
3141
|
_provider_rc=$?
|
|
3136
3142
|
fi
|
|
3137
3143
|
;;
|
|
3144
|
+
*)
|
|
3145
|
+
# v8.2.0 TIMEOUT SEAM. Any provider exposing provider_invoke_argv can
|
|
3146
|
+
# cast a real council vote instead of falling straight to the
|
|
3147
|
+
# heuristic. argv is a REAL command, so `timeout` bounds it exactly as
|
|
3148
|
+
# it bounds the named arms above (providers/claude.sh:321).
|
|
3149
|
+
#
|
|
3150
|
+
# SEMANTICS PRESERVED: _provider_rc is captured the same way, so the
|
|
3151
|
+
# bash-F4 safe default below still forces a conservative REJECT on a
|
|
3152
|
+
# timeout kill (124/137/143). An empty verdict falls to
|
|
3153
|
+
# council_heuristic_review, identical to a missing CLI today.
|
|
3154
|
+
if type provider_invoke_argv >/dev/null 2>&1; then
|
|
3155
|
+
provider_invoke_argv fast "$prompt"
|
|
3156
|
+
# caveman HARD-SUPPRESS: this vote is parsed for "VOTE:".
|
|
3157
|
+
verdict=$(timeout "${LOKI_COUNCIL_REVIEW_TIMEOUT:-600}" \
|
|
3158
|
+
env CAVEMAN_DEFAULT_MODE=off \
|
|
3159
|
+
"${_LOKI_INVOKE_ARGV[@]+"${_LOKI_INVOKE_ARGV[@]}"}" 2>/dev/null)
|
|
3160
|
+
_provider_rc=$?
|
|
3161
|
+
fi
|
|
3162
|
+
;;
|
|
3138
3163
|
esac
|
|
3139
3164
|
|
|
3140
3165
|
# bash-F4 (WAVE10 SAFE-DEFAULT): a provider timeout (124, incl. 128+SIGTERM
|
|
@@ -3303,6 +3328,22 @@ REASON: your reasoning"
|
|
|
3303
3328
|
verdict=$(timeout "${LOKI_COUNCIL_REVIEW_TIMEOUT:-600}" aider --message "$prompt" --yes-always --no-auto-commits --no-git 2>/dev/null)
|
|
3304
3329
|
fi
|
|
3305
3330
|
;;
|
|
3331
|
+
*)
|
|
3332
|
+
# v8.2.0 TIMEOUT SEAM (contrarian / devil's-advocate vote). Same
|
|
3333
|
+
# rationale as the member vote: a real argv keeps the `timeout`
|
|
3334
|
+
# bound that a shell function would silently remove.
|
|
3335
|
+
#
|
|
3336
|
+
# SEMANTICS PRESERVED: this path tracks no _provider_rc by design --
|
|
3337
|
+
# an empty verdict (timeout or failure) already routes to the
|
|
3338
|
+
# conservative REJECT fallback immediately below.
|
|
3339
|
+
if type provider_invoke_argv >/dev/null 2>&1; then
|
|
3340
|
+
provider_invoke_argv fast "$prompt"
|
|
3341
|
+
# caveman HARD-SUPPRESS: parsed for "VOTE:".
|
|
3342
|
+
verdict=$(timeout "${LOKI_COUNCIL_REVIEW_TIMEOUT:-600}" \
|
|
3343
|
+
env CAVEMAN_DEFAULT_MODE=off \
|
|
3344
|
+
"${_LOKI_INVOKE_ARGV[@]+"${_LOKI_INVOKE_ARGV[@]}"}" 2>/dev/null)
|
|
3345
|
+
fi
|
|
3346
|
+
;;
|
|
3306
3347
|
esac
|
|
3307
3348
|
|
|
3308
3349
|
if [ -z "$verdict" ]; then
|
package/autonomy/council-v2.sh
CHANGED
|
@@ -82,6 +82,9 @@ council_v2_vote() {
|
|
|
82
82
|
local first=true
|
|
83
83
|
local approve_count=0
|
|
84
84
|
local reject_count=0
|
|
85
|
+
# Reviewers we could not obtain a verdict from. Never folded into
|
|
86
|
+
# reject_count -- see the TRUST comment in the tally below.
|
|
87
|
+
local inconclusive_count=0
|
|
85
88
|
for vote_file in "${vote_files[@]}"; do
|
|
86
89
|
if [ -f "$vote_file" ]; then
|
|
87
90
|
local vote_content
|
|
@@ -95,16 +98,40 @@ council_v2_vote() {
|
|
|
95
98
|
|
|
96
99
|
local verdict
|
|
97
100
|
verdict=$(echo "$vote_content" | python3 -c "import sys,json; print(json.load(sys.stdin).get('verdict','').upper())" 2>/dev/null || echo "UNKNOWN")
|
|
101
|
+
# TRUST: a reviewer that could not be REACHED did not vote REJECT.
|
|
102
|
+
#
|
|
103
|
+
# Every provider arm below used to substitute a literal
|
|
104
|
+
# {"verdict":"REJECT","reasoning":"review execution failed"} on any
|
|
105
|
+
# miss -- CLI absent, timeout, transient error, unparseable output --
|
|
106
|
+
# and this tally counted "anything not APPROVE" as a rejection. The
|
|
107
|
+
# engine therefore fabricated a reviewer vote the model never gave,
|
|
108
|
+
# and the run was BLOCKED by it.
|
|
109
|
+
#
|
|
110
|
+
# That is the same defect class as a receipt attesting to a diff stat
|
|
111
|
+
# it did not measure: the artifact claims a fact nobody established.
|
|
112
|
+
# On a weak model that formats poorly it is worse still -- low format
|
|
113
|
+
# compliance becomes a permanent BLOCK that reads to the user as
|
|
114
|
+
# "Loki is broken" rather than "your model could not answer".
|
|
115
|
+
#
|
|
116
|
+
# INCONCLUSIVE is counted separately and never as a rejection.
|
|
98
117
|
if [ "$verdict" = "APPROVE" ]; then
|
|
99
118
|
((approve_count++))
|
|
100
|
-
|
|
119
|
+
elif [ "$verdict" = "REJECT" ]; then
|
|
101
120
|
((reject_count++))
|
|
121
|
+
else
|
|
122
|
+
((inconclusive_count++))
|
|
102
123
|
fi
|
|
103
124
|
fi
|
|
104
125
|
done
|
|
105
126
|
votes_json="$votes_json]"
|
|
106
127
|
|
|
107
|
-
|
|
128
|
+
if [ "$inconclusive_count" -gt 0 ]; then
|
|
129
|
+
log_warn "Blind review results: $approve_count APPROVE / $reject_count REJECT / $inconclusive_count INCONCLUSIVE"
|
|
130
|
+
log_warn " $inconclusive_count reviewer(s) produced no verdict (CLI missing, timeout, or unparseable output)."
|
|
131
|
+
log_warn " These are NOT rejections. The council decides on the verdicts it actually obtained."
|
|
132
|
+
else
|
|
133
|
+
log_info "Blind review results: $approve_count APPROVE / $reject_count REJECT"
|
|
134
|
+
fi
|
|
108
135
|
|
|
109
136
|
# Step 4: Sycophancy detection
|
|
110
137
|
local sycophancy_score
|
|
@@ -137,8 +164,13 @@ print('{:.3f}'.format(detect_sycophancy(votes)))
|
|
|
137
164
|
|
|
138
165
|
if [ -f "$da_vote" ]; then
|
|
139
166
|
local da_verdict
|
|
140
|
-
|
|
141
|
-
|
|
167
|
+
# Same trust rule as the main tally: an unparseable devil's
|
|
168
|
+
# advocate did not vote REJECT. Defaulting to REJECT here would
|
|
169
|
+
# silently overturn a unanimous APPROVE on a transient failure.
|
|
170
|
+
da_verdict=$(cat "$da_vote" | python3 -c "import sys,json; print(json.load(sys.stdin).get('verdict','').upper())" 2>/dev/null || echo "INCONCLUSIVE")
|
|
171
|
+
if [ "$da_verdict" = "INCONCLUSIVE" ]; then
|
|
172
|
+
log_warn "Devil's advocate produced no verdict -- unanimous approval left UNCHANGED (not overturned)"
|
|
173
|
+
elif [ "$da_verdict" = "REJECT" ]; then
|
|
142
174
|
log_warn "Devil's advocate REJECTED unanimous approval"
|
|
143
175
|
approve_count=$((approve_count - 1))
|
|
144
176
|
reject_count=$((reject_count + 1))
|
|
@@ -374,42 +406,64 @@ except Exception:
|
|
|
374
406
|
fi
|
|
375
407
|
# Fall through to the plain text call (+ sed-carve) on any miss.
|
|
376
408
|
if [ -z "$result" ]; then
|
|
377
|
-
result=$(echo "$full_prompt" | CAVEMAN_DEFAULT_MODE=off claude "${_c2_argv[@]}" -p 2>/dev/null || echo '{"verdict":"
|
|
409
|
+
result=$(echo "$full_prompt" | CAVEMAN_DEFAULT_MODE=off claude "${_c2_argv[@]}" -p 2>/dev/null || echo '{"verdict":"INCONCLUSIVE","reasoning":"review execution failed (no verdict obtained; NOT a rejection)","issues":[]}')
|
|
378
410
|
fi
|
|
379
411
|
else
|
|
380
|
-
result='{"verdict":"
|
|
412
|
+
result='{"verdict":"INCONCLUSIVE","reasoning":"reviewer CLI unavailable (no verdict obtained; NOT a rejection)","issues":[]}'
|
|
381
413
|
fi
|
|
382
414
|
;;
|
|
383
415
|
codex)
|
|
384
416
|
if command -v codex &>/dev/null; then
|
|
385
|
-
result=$(codex exec -q "$full_prompt" 2>/dev/null || echo '{"verdict":"
|
|
417
|
+
result=$(codex exec -q "$full_prompt" 2>/dev/null || echo '{"verdict":"INCONCLUSIVE","reasoning":"review execution failed (no verdict obtained; NOT a rejection)","issues":[]}')
|
|
386
418
|
else
|
|
387
|
-
result='{"verdict":"
|
|
419
|
+
result='{"verdict":"INCONCLUSIVE","reasoning":"reviewer CLI unavailable (no verdict obtained; NOT a rejection)","issues":[]}'
|
|
388
420
|
fi
|
|
389
421
|
;;
|
|
390
422
|
gemini)
|
|
391
423
|
if command -v gemini &>/dev/null; then
|
|
392
|
-
result=$(echo "$full_prompt" | gemini 2>/dev/null || echo '{"verdict":"
|
|
424
|
+
result=$(echo "$full_prompt" | gemini 2>/dev/null || echo '{"verdict":"INCONCLUSIVE","reasoning":"review execution failed (no verdict obtained; NOT a rejection)","issues":[]}')
|
|
393
425
|
else
|
|
394
|
-
result='{"verdict":"
|
|
426
|
+
result='{"verdict":"INCONCLUSIVE","reasoning":"reviewer CLI unavailable (no verdict obtained; NOT a rejection)","issues":[]}'
|
|
395
427
|
fi
|
|
396
428
|
;;
|
|
397
429
|
cline)
|
|
398
430
|
if command -v cline &>/dev/null; then
|
|
399
|
-
result=$(cline -y "$full_prompt" 2>/dev/null || echo '{"verdict":"
|
|
431
|
+
result=$(cline -y "$full_prompt" 2>/dev/null || echo '{"verdict":"INCONCLUSIVE","reasoning":"review execution failed (no verdict obtained; NOT a rejection)","issues":[]}')
|
|
400
432
|
else
|
|
401
|
-
result='{"verdict":"
|
|
433
|
+
result='{"verdict":"INCONCLUSIVE","reasoning":"reviewer CLI unavailable (no verdict obtained; NOT a rejection)","issues":[]}'
|
|
402
434
|
fi
|
|
403
435
|
;;
|
|
404
436
|
aider)
|
|
405
437
|
if command -v aider &>/dev/null; then
|
|
406
|
-
result=$(aider --message "$full_prompt" --yes-always --no-auto-commits --no-git 2>/dev/null || echo '{"verdict":"
|
|
438
|
+
result=$(aider --message "$full_prompt" --yes-always --no-auto-commits --no-git 2>/dev/null || echo '{"verdict":"INCONCLUSIVE","reasoning":"review execution failed (no verdict obtained; NOT a rejection)","issues":[]}')
|
|
407
439
|
else
|
|
408
|
-
result='{"verdict":"
|
|
440
|
+
result='{"verdict":"INCONCLUSIVE","reasoning":"reviewer CLI unavailable (no verdict obtained; NOT a rejection)","issues":[]}'
|
|
409
441
|
fi
|
|
410
442
|
;;
|
|
411
443
|
*)
|
|
412
|
-
|
|
444
|
+
# v8.2.0 TIMEOUT SEAM. A provider exposing provider_invoke_argv gets
|
|
445
|
+
# a real reviewer instead of an automatic INCONCLUSIVE. The argv is a
|
|
446
|
+
# REAL command so `timeout` genuinely bounds it (providers/claude.sh:321);
|
|
447
|
+
# routing this through a shell function would silently drop the bound.
|
|
448
|
+
#
|
|
449
|
+
# VERDICT SEMANTICS UNCHANGED: on empty output, non-zero exit, or a
|
|
450
|
+
# timeout kill (124/137/143) we fall back to the SAME INCONCLUSIVE
|
|
451
|
+
# string used below -- never REJECT. A judge that produced no
|
|
452
|
+
# judgement has not rejected anything (v8.1.0 TRUST-3).
|
|
453
|
+
result=''
|
|
454
|
+
if type provider_invoke_argv >/dev/null 2>&1; then
|
|
455
|
+
local _c2_seam_rc=0
|
|
456
|
+
provider_invoke_argv fast "$full_prompt"
|
|
457
|
+
# caveman HARD-SUPPRESS: this verdict is parsed for the JSON
|
|
458
|
+
# "verdict" field; compression would reword it.
|
|
459
|
+
result="$(timeout "${LOKI_SDK_REVIEW_TIMEOUT:-180}" \
|
|
460
|
+
env CAVEMAN_DEFAULT_MODE=off \
|
|
461
|
+
"${_LOKI_INVOKE_ARGV[@]+"${_LOKI_INVOKE_ARGV[@]}"}" 2>/dev/null)" || _c2_seam_rc=$?
|
|
462
|
+
[ "$_c2_seam_rc" -ne 0 ] && result=''
|
|
463
|
+
fi
|
|
464
|
+
if [ -z "$result" ]; then
|
|
465
|
+
result='{"verdict":"INCONCLUSIVE","reasoning":"review not supported for this provider (no verdict obtained; NOT a rejection)","issues":[]}'
|
|
466
|
+
fi
|
|
413
467
|
;;
|
|
414
468
|
esac
|
|
415
469
|
|
|
@@ -420,8 +474,52 @@ except Exception:
|
|
|
420
474
|
# Try removing markdown fencing
|
|
421
475
|
extracted=$(echo "$result" | sed 's/^```json//;s/^```//' | sed -n '/^{/,/^}/p' | head -50)
|
|
422
476
|
fi
|
|
477
|
+
# STRUCTURE-TOLERANT RECOVERY (v8.2.0).
|
|
478
|
+
#
|
|
479
|
+
# A strict JSON carve is the single most model-sensitive contract in the
|
|
480
|
+
# engine: schema adherence is exactly what varies most across models, while
|
|
481
|
+
# every coding model can state a verdict in prose. Measured elsewhere (Forge
|
|
482
|
+
# replication): scaffolding drove tool-call errors 42 -> 0 while
|
|
483
|
+
# advanced-reasoning accuracy stayed flat -- i.e. a harness CAN rescue
|
|
484
|
+
# format compliance but cannot manufacture judgment. So recovering a verdict
|
|
485
|
+
# the model genuinely expressed is legitimate; inventing one is not.
|
|
486
|
+
#
|
|
487
|
+
# This runs ONLY when the JSON carve produced nothing, and it accepts a
|
|
488
|
+
# verdict only when the model stated it UNAMBIGUOUSLY (exactly one of
|
|
489
|
+
# APPROVE/REJECT appears as a standalone word). A response mentioning both,
|
|
490
|
+
# or neither, stays INCONCLUSIVE -- never guessed.
|
|
491
|
+
if [ -z "$extracted" ] && [ -n "${result:-}" ]; then
|
|
492
|
+
local _recovered
|
|
493
|
+
_recovered="$(printf '%s' "$result" | _LOKI_RAW="$result" python3 -c '
|
|
494
|
+
import os, re, sys, json
|
|
495
|
+
raw = os.environ.get("_LOKI_RAW", "")
|
|
496
|
+
# Standalone words only: "APPROVE" not "approved-by", and not inside a URL.
|
|
497
|
+
approve = len(re.findall(r"(?<![A-Za-z0-9_-])APPROVE(?![A-Za-z0-9_-])", raw, re.I))
|
|
498
|
+
reject = len(re.findall(r"(?<![A-Za-z0-9_-])REJECT(?![A-Za-z0-9_-])", raw, re.I))
|
|
499
|
+
if approve and not reject:
|
|
500
|
+
v = "APPROVE"
|
|
501
|
+
elif reject and not approve:
|
|
502
|
+
v = "REJECT"
|
|
503
|
+
else:
|
|
504
|
+
sys.exit(1) # ambiguous or absent -> stay inconclusive
|
|
505
|
+
print(json.dumps({
|
|
506
|
+
"verdict": v,
|
|
507
|
+
"reasoning": "recovered from unstructured output (model stated %s in prose)" % v,
|
|
508
|
+
"issues": [],
|
|
509
|
+
"recovered": True,
|
|
510
|
+
}))
|
|
511
|
+
' 2>/dev/null)" || _recovered=""
|
|
512
|
+
if [ -n "$_recovered" ]; then
|
|
513
|
+
extracted="$_recovered"
|
|
514
|
+
fi
|
|
515
|
+
fi
|
|
516
|
+
|
|
423
517
|
if [ -z "$extracted" ]; then
|
|
424
|
-
|
|
518
|
+
# The weak-model case, and the reason this matters most: a model whose
|
|
519
|
+
# prose could not be carved into JSON did not vote REJECT. Recording one
|
|
520
|
+
# here would turn "your model formats poorly" into "your code was
|
|
521
|
+
# rejected" -- a block the user cannot act on and did not earn.
|
|
522
|
+
extracted='{"verdict":"INCONCLUSIVE","reasoning":"failed to parse review output (no verdict obtained; NOT a rejection)","issues":[]}'
|
|
425
523
|
fi
|
|
426
524
|
|
|
427
525
|
echo "$extracted" > "$output_file"
|
package/autonomy/grill.sh
CHANGED
|
@@ -158,6 +158,12 @@ grill_check_provider() {
|
|
|
158
158
|
fi
|
|
159
159
|
;;
|
|
160
160
|
*)
|
|
161
|
+
# v8.2.0: capability, not identity. Must stay in lockstep with the
|
|
162
|
+
# matching arm in grill_invoke_provider -- this gate runs FIRST, so
|
|
163
|
+
# rejecting here would make that arm dead code.
|
|
164
|
+
if type provider_invoke_argv >/dev/null 2>&1; then
|
|
165
|
+
return 0
|
|
166
|
+
fi
|
|
161
167
|
_grill_err "grill currently supports the claude and codex providers (got: $provider)"
|
|
162
168
|
return $GRILL_EXIT_ERROR
|
|
163
169
|
;;
|
|
@@ -278,6 +284,32 @@ grill_invoke_provider() {
|
|
|
278
284
|
return 0
|
|
279
285
|
;;
|
|
280
286
|
*)
|
|
287
|
+
# v8.2.0 TIMEOUT SEAM. Previously any other provider was a hard
|
|
288
|
+
# error. A provider exposing provider_invoke_argv builds a REAL argv
|
|
289
|
+
# (not a shell function), so _grill_with_timeout still bounds it --
|
|
290
|
+
# the whole reason the seam exists (providers/claude.sh:321).
|
|
291
|
+
# Nothing to preserve on this arm, so nothing can regress: it
|
|
292
|
+
# produced no output at all before.
|
|
293
|
+
#
|
|
294
|
+
# Deliberate: no --disallowedTools. That flag is claude-specific with
|
|
295
|
+
# no portable equivalent, so a seam-provider grill CAN write to the
|
|
296
|
+
# tree. Accepted; the alternative is the hard error below.
|
|
297
|
+
if type provider_invoke_argv >/dev/null 2>&1; then
|
|
298
|
+
local out
|
|
299
|
+
provider_invoke_argv fast "$prompt"
|
|
300
|
+
# env CAVEMAN_DEFAULT_MODE=off: grill output is parsed downstream
|
|
301
|
+
# and caveman compression would reword the questions. `env` (not a
|
|
302
|
+
# bare prefix) because _grill_with_timeout execs its first token.
|
|
303
|
+
out="$(_grill_with_timeout "${LOKI_GRILL_TIMEOUT:-180}" \
|
|
304
|
+
env CAVEMAN_DEFAULT_MODE=off \
|
|
305
|
+
"${_LOKI_INVOKE_ARGV[@]+"${_LOKI_INVOKE_ARGV[@]}"}" 2>/dev/null)"
|
|
306
|
+
if [ -z "$out" ]; then
|
|
307
|
+
_grill_err "provider returned no output (timeout or invocation error)"
|
|
308
|
+
return $GRILL_EXIT_ERROR
|
|
309
|
+
fi
|
|
310
|
+
printf '%s\n' "$out"
|
|
311
|
+
return 0
|
|
312
|
+
fi
|
|
281
313
|
_grill_err "grill currently supports the claude and codex providers (got: $provider)"
|
|
282
314
|
return $GRILL_EXIT_ERROR
|
|
283
315
|
;;
|
|
@@ -159,12 +159,24 @@ except Exception:
|
|
|
159
159
|
return 0
|
|
160
160
|
}
|
|
161
161
|
|
|
162
|
-
# Decide whether model verification can be attempted.
|
|
163
|
-
#
|
|
164
|
-
#
|
|
162
|
+
# Decide whether model verification can be attempted.
|
|
163
|
+
#
|
|
164
|
+
# v8.2.0: this used to require LOKI_PROVIDER=claude specifically, so a user on
|
|
165
|
+
# codex/opencode/cline/aider silently lost done-recognition. That is a
|
|
166
|
+
# CAPABILITY question, not an identity question -- the real requirement is "can
|
|
167
|
+
# we reach a model AND bound the call with a timeout". Any provider exposing the
|
|
168
|
+
# argv seam (provider_invoke_argv, see providers/claude.sh) satisfies both.
|
|
169
|
+
#
|
|
170
|
+
# Falls back to the historical claude-binary check when the seam is absent, so
|
|
171
|
+
# nothing regresses for existing installs.
|
|
165
172
|
_loki_done_recog_provider_ok() {
|
|
166
|
-
[ "${LOKI_PROVIDER:-claude}" = "claude" ] || return 1
|
|
167
173
|
[ "${PROVIDER_DEGRADED:-false}" != "true" ] || return 1
|
|
174
|
+
# Preferred: a provider that can build a timeout-able argv.
|
|
175
|
+
if type provider_invoke_argv >/dev/null 2>&1; then
|
|
176
|
+
return 0
|
|
177
|
+
fi
|
|
178
|
+
# Legacy path: claude binary present.
|
|
179
|
+
[ "${LOKI_PROVIDER:-claude}" = "claude" ] || return 1
|
|
168
180
|
command -v claude >/dev/null 2>&1 || return 1
|
|
169
181
|
return 0
|
|
170
182
|
}
|