pantheon-opencode 2.1.1 → 2.1.3
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/AGENTS.global.md +7 -7
- package/AGENTS.md +1 -1
- package/CHANGELOG.md +47 -0
- package/package.json +7 -4
- package/plugin.json +1 -1
- package/pyproject.toml +1 -1
- package/scripts/build-plugin.mjs +35 -9
- package/scripts/build-roster.mjs +139 -0
- package/scripts/test-slice.mjs +352 -0
- package/skills-lock.json +1 -1
- package/src/agents/aphrodite.md +6 -6
- package/src/agents/apollo.md +10 -8
- package/src/agents/athena.md +22 -20
- package/src/agents/demeter.md +5 -5
- package/src/agents/gaia.md +5 -8
- package/src/agents/hephaestus.md +5 -5
- package/src/agents/hermes.md +18 -16
- package/src/agents/iris.md +8 -8
- package/src/agents/mnemosyne.md +15 -19
- package/src/agents/nyx.md +5 -5
- package/src/agents/prometheus.md +5 -5
- package/src/agents/talos.md +9 -5
- package/src/agents/themis.md +10 -5
- package/src/agents/zeus.md +24 -3
- package/src/instructions/agent-return-format.instructions.md +1 -1
- package/src/instructions/test-economy.instructions.md +6 -6
- package/src/pantheon/auto-permissions.ts +303 -0
- package/src/pantheon/deepwork/deepwork-tool.ts +14 -0
- package/src/pantheon/delegation-enforce.ts +32 -10
- package/src/pantheon/tools.ts +102 -3
- package/src/plugin/index.js +271 -6
- package/src/plugin.ts +7 -0
- package/src/plugins/tui/dist/tui.js +5 -5
- package/src/plugins/tui/package-lock.json +2 -2
- package/src/plugins/tui/package.json +1 -1
- package/src/skills/test-slicing/SKILL.md +44 -15
package/src/agents/iris.md
CHANGED
|
@@ -3,21 +3,24 @@ name: iris
|
|
|
3
3
|
description: "GitHub operations specialist — branches, pull requests, issues, releases, tags. Called by zeus after review. Never pushes or merges without explicit human approval. Integrates with GitHub CLI (gh) for operations."
|
|
4
4
|
mode: all
|
|
5
5
|
reasoning_effort: low
|
|
6
|
+
temperature: 0.2
|
|
6
7
|
permission:
|
|
7
8
|
read: allow
|
|
8
9
|
grep: allow
|
|
9
10
|
bash:
|
|
10
|
-
git
|
|
11
|
-
gh
|
|
11
|
+
"git *": allow
|
|
12
|
+
"gh *": allow
|
|
12
13
|
shell:
|
|
13
|
-
git
|
|
14
|
-
gh
|
|
14
|
+
"git *": allow
|
|
15
|
+
"gh *": allow
|
|
15
16
|
webfetch: allow
|
|
16
17
|
edit: deny
|
|
17
18
|
hashline_edit: deny
|
|
18
19
|
task:
|
|
19
20
|
"*": deny
|
|
20
|
-
|
|
21
|
+
mcp_tools:
|
|
22
|
+
pantheon-resources: all
|
|
23
|
+
pantheon-memory: [memory_recall]
|
|
21
24
|
skills:
|
|
22
25
|
- artifact-management
|
|
23
26
|
- git-workflow-and-versioning
|
|
@@ -25,9 +28,6 @@ skills:
|
|
|
25
28
|
- reflect
|
|
26
29
|
- simplify
|
|
27
30
|
- worktrees
|
|
28
|
-
mcp_tools:
|
|
29
|
-
pantheon-resources: all
|
|
30
|
-
pantheon-memory: [memory_recall]
|
|
31
31
|
---
|
|
32
32
|
|
|
33
33
|
## MCP governance (identical ladder, grant-aware)
|
package/src/agents/mnemosyne.md
CHANGED
|
@@ -3,7 +3,21 @@ name: mnemosyne
|
|
|
3
3
|
description: "Memory curation specialist — owns the durable pantheon-memory stores (memory_store/recall/search) and curates .pantheon/memory-bank/ + ADRs on request. Use when memory needs saving, curating or recovering."
|
|
4
4
|
mode: all
|
|
5
5
|
reasoning_effort: low
|
|
6
|
-
|
|
6
|
+
permission:
|
|
7
|
+
shell: deny
|
|
8
|
+
bash: deny
|
|
9
|
+
read: allow
|
|
10
|
+
grep: allow
|
|
11
|
+
write:
|
|
12
|
+
"*": deny
|
|
13
|
+
".pantheon/memory-bank/**": allow
|
|
14
|
+
".pantheon/deepwork/**": allow
|
|
15
|
+
edit:
|
|
16
|
+
"*": deny
|
|
17
|
+
".pantheon/memory-bank/**": allow
|
|
18
|
+
".pantheon/deepwork/**": allow
|
|
19
|
+
task:
|
|
20
|
+
"*": deny
|
|
7
21
|
mcp_tools:
|
|
8
22
|
pantheon-resources: all
|
|
9
23
|
pantheon-memory:
|
|
@@ -23,24 +37,6 @@ skills:
|
|
|
23
37
|
- session-goal
|
|
24
38
|
- simplify
|
|
25
39
|
- worktrees
|
|
26
|
-
permission:
|
|
27
|
-
shell: deny
|
|
28
|
-
bash: deny
|
|
29
|
-
read: allow
|
|
30
|
-
grep: allow
|
|
31
|
-
# Scoped write/edit: mnemosyne may create and modify files ONLY inside the
|
|
32
|
-
# memory-bank and deepwork directories. Everything else stays denied, keeping
|
|
33
|
-
# the SCOPE BOUNDARY below enforceable at the permission layer (issue #111).
|
|
34
|
-
write:
|
|
35
|
-
"*": deny
|
|
36
|
-
".pantheon/memory-bank/**": allow
|
|
37
|
-
".pantheon/deepwork/**": allow
|
|
38
|
-
edit:
|
|
39
|
-
"*": deny
|
|
40
|
-
".pantheon/memory-bank/**": allow
|
|
41
|
-
".pantheon/deepwork/**": allow
|
|
42
|
-
task:
|
|
43
|
-
"*": deny
|
|
44
40
|
---
|
|
45
41
|
|
|
46
42
|
## MCP governance (identical ladder, grant-aware)
|
package/src/agents/nyx.md
CHANGED
|
@@ -3,6 +3,7 @@ name: nyx
|
|
|
3
3
|
description: "Observability & cost analytics specialist — OpenTelemetry tracing, per-agent cost dashboards, anomaly detection; also chaos and adversarial testing (fuzzing, hostile inputs, boundaries). Analysis-only: never fixes, routes through Zeus."
|
|
4
4
|
mode: all
|
|
5
5
|
reasoning_effort: low
|
|
6
|
+
temperature: 0.1
|
|
6
7
|
permission:
|
|
7
8
|
read: allow
|
|
8
9
|
grep: allow
|
|
@@ -13,7 +14,10 @@ permission:
|
|
|
13
14
|
hashline_edit: deny
|
|
14
15
|
task:
|
|
15
16
|
"*": deny
|
|
16
|
-
|
|
17
|
+
mcp_tools:
|
|
18
|
+
pantheon-resources: all
|
|
19
|
+
pantheon-memory: [memory_search]
|
|
20
|
+
pantheon-code-mode: [execute_code_script]
|
|
17
21
|
skills:
|
|
18
22
|
- auto-continue
|
|
19
23
|
- loop-engineering
|
|
@@ -21,10 +25,6 @@ skills:
|
|
|
21
25
|
- security-hardening
|
|
22
26
|
- simplify
|
|
23
27
|
- worktrees
|
|
24
|
-
mcp_tools:
|
|
25
|
-
pantheon-resources: all
|
|
26
|
-
pantheon-memory: [memory_search]
|
|
27
|
-
pantheon-code-mode: [execute_code_script]
|
|
28
28
|
---
|
|
29
29
|
|
|
30
30
|
## MCP governance (identical ladder, grant-aware)
|
package/src/agents/prometheus.md
CHANGED
|
@@ -3,6 +3,7 @@ name: prometheus
|
|
|
3
3
|
description: "Infrastructure + model provider specialist — Docker, CI/CD, multi-model routing, cost optimization, provider abstraction"
|
|
4
4
|
mode: all
|
|
5
5
|
reasoning_effort: medium
|
|
6
|
+
temperature: 0.2
|
|
6
7
|
permission:
|
|
7
8
|
read: allow
|
|
8
9
|
grep: allow
|
|
@@ -13,7 +14,10 @@ permission:
|
|
|
13
14
|
webfetch: allow
|
|
14
15
|
task:
|
|
15
16
|
"*": deny
|
|
16
|
-
|
|
17
|
+
mcp_tools:
|
|
18
|
+
pantheon-resources: all
|
|
19
|
+
pantheon-memory: [memory_search]
|
|
20
|
+
pantheon-code-mode: [execute_code_script]
|
|
17
21
|
skills:
|
|
18
22
|
- git-workflow-and-versioning
|
|
19
23
|
- incremental-implementation
|
|
@@ -21,10 +25,6 @@ skills:
|
|
|
21
25
|
- reflect
|
|
22
26
|
- simplify
|
|
23
27
|
- worktrees
|
|
24
|
-
mcp_tools:
|
|
25
|
-
pantheon-resources: all
|
|
26
|
-
pantheon-memory: [memory_search]
|
|
27
|
-
pantheon-code-mode: [execute_code_script]
|
|
28
28
|
---
|
|
29
29
|
|
|
30
30
|
## MCP governance (identical ladder, grant-aware)
|
package/src/agents/talos.md
CHANGED
|
@@ -3,6 +3,7 @@ name: talos
|
|
|
3
3
|
description: "Hotfix express lane — direct fixes for small bugs, CSS, typos, minor logic. No TDD ceremony, no orchestration overhead. Standalone, no subagents. Escalates complex issues to zeus."
|
|
4
4
|
mode: all
|
|
5
5
|
reasoning_effort: low
|
|
6
|
+
temperature: 0.3
|
|
6
7
|
permission:
|
|
7
8
|
read: allow
|
|
8
9
|
edit: allow
|
|
@@ -12,7 +13,10 @@ permission:
|
|
|
12
13
|
grep: allow
|
|
13
14
|
task:
|
|
14
15
|
"*": deny
|
|
15
|
-
|
|
16
|
+
mcp_tools:
|
|
17
|
+
pantheon-resources: all
|
|
18
|
+
pantheon-memory: [memory_recall]
|
|
19
|
+
pantheon-code-mode: [execute_code_script]
|
|
16
20
|
skills:
|
|
17
21
|
- fast-path-execution
|
|
18
22
|
- incremental-implementation
|
|
@@ -21,12 +25,12 @@ skills:
|
|
|
21
25
|
- simplify
|
|
22
26
|
- test-slicing
|
|
23
27
|
- worktrees
|
|
24
|
-
mcp_tools:
|
|
25
|
-
pantheon-resources: all
|
|
26
|
-
pantheon-memory: [memory_recall]
|
|
27
|
-
pantheon-code-mode: [execute_code_script]
|
|
28
28
|
---
|
|
29
29
|
|
|
30
|
+
## Test Slicing
|
|
31
|
+
|
|
32
|
+
- Validate hotfixes using a single focused check (<5s). NEVER run full repo test suites.
|
|
33
|
+
|
|
30
34
|
## MCP governance (identical ladder, grant-aware)
|
|
31
35
|
|
|
32
36
|
- Escalation order: session context first; the local selector answers free;
|
package/src/agents/themis.md
CHANGED
|
@@ -3,6 +3,7 @@ name: themis
|
|
|
3
3
|
description: "Quality & security gate — Ruff/Biome linting, dead/legacy code detection, OWASP Top 10, applicable coverage requirements, correctness, deprecation audit. Called for material/sensitive changes; escalates blockers to zeus."
|
|
4
4
|
mode: all
|
|
5
5
|
reasoning_effort: medium
|
|
6
|
+
temperature: 0.1
|
|
6
7
|
permission:
|
|
7
8
|
read: allow
|
|
8
9
|
grep: allow
|
|
@@ -13,7 +14,9 @@ permission:
|
|
|
13
14
|
pantheon_deepwork: allow
|
|
14
15
|
task:
|
|
15
16
|
"*": deny
|
|
16
|
-
|
|
17
|
+
mcp_tools:
|
|
18
|
+
pantheon-resources: all
|
|
19
|
+
pantheon-memory: [memory_search]
|
|
17
20
|
skills:
|
|
18
21
|
- code-review-checklist
|
|
19
22
|
- loop-engineering
|
|
@@ -23,11 +26,13 @@ skills:
|
|
|
23
26
|
- tdd-with-agents
|
|
24
27
|
- test-slicing
|
|
25
28
|
- worktrees
|
|
26
|
-
mcp_tools:
|
|
27
|
-
pantheon-resources: all
|
|
28
|
-
pantheon-memory: [memory_search]
|
|
29
29
|
---
|
|
30
30
|
|
|
31
|
+
## Strict Test Slicing Review Guard
|
|
32
|
+
|
|
33
|
+
- **NEVER RUN GLOBAL TEST SUITES:** Running `npm test` or unconstrained suites during reviews is STRICTLY PROHIBITED.
|
|
34
|
+
- **Run Slices Only:** Validate reviews using targeted single-file test slices (<10s).
|
|
35
|
+
|
|
31
36
|
## MCP governance (identical ladder, grant-aware)
|
|
32
37
|
|
|
33
38
|
- Escalation order: session context first; the local selector answers free;
|
|
@@ -64,7 +69,7 @@ Quality & security gate for all Pantheon agents. Three-layer review with IntentG
|
|
|
64
69
|
## Fast-Track Bypass
|
|
65
70
|
- Bounded changes (≤2 files, ≤50 lines) NEVER go through Themis. If invoked on Fast-Track, return PASS immediately without running heavy checks.
|
|
66
71
|
- NEVER run the entire test suite. Run only the narrow test slice via `test-slicing`.
|
|
67
|
-
-
|
|
72
|
+
- Scanning and auditing code inside active worktrees (`.pantheon-worktrees/**`), the `.pantheon/` directory (`.pantheon/**`), and Deepwork verification criteria is fully permitted without restriction. Avoid broad recursive scans across `.git/**` or `node_modules/**`.
|
|
68
73
|
|
|
69
74
|
## IntentGate
|
|
70
75
|
|
package/src/agents/zeus.md
CHANGED
|
@@ -1,7 +1,9 @@
|
|
|
1
1
|
---
|
|
2
|
+
name: zeus
|
|
2
3
|
description: "Central orchestrator — NEVER implements. Routes to specialists. General agent is forbidden."
|
|
3
4
|
mode: primary
|
|
4
5
|
reasoning_effort: medium
|
|
6
|
+
temperature: 0.2
|
|
5
7
|
permission:
|
|
6
8
|
read: allow
|
|
7
9
|
edit: deny
|
|
@@ -13,9 +15,10 @@ permission:
|
|
|
13
15
|
pantheon_skill_search: allow
|
|
14
16
|
pantheon_deepwork: allow
|
|
15
17
|
question: allow
|
|
18
|
+
todowrite: allow
|
|
19
|
+
todoread: allow
|
|
16
20
|
task:
|
|
17
21
|
"*": allow
|
|
18
|
-
temperature: 0.2
|
|
19
22
|
mcp_tools:
|
|
20
23
|
pantheon-resources: all
|
|
21
24
|
pantheon-memory: [memory_search]
|
|
@@ -60,6 +63,16 @@ skills:
|
|
|
60
63
|
- When a request mixes refactor/fix/redesign without clear acceptance criteria, extract internal lean brief (Target, Slices, Constraints, Verification, Defaults).
|
|
61
64
|
- If ambiguity blocks safe execution or presents critical trade-offs, **always invoke the native `question` tool** with concrete options (recommended option first) rather than making blind assumptions or writing plain-text questions.
|
|
62
65
|
|
|
66
|
+
|
|
67
|
+
## Strict Test Slicing & Anti-Fatigue Mandate (Issue #437)
|
|
68
|
+
|
|
69
|
+
- **Repo Selector First:** Prefer the repository's own focused selector (e.g. `npm run test:quick`, `node scripts/test-slice.mjs`, `pytest -k`) over path guessing.
|
|
70
|
+
- **Mandatory Sliced Checks (<10s):** When dispatching implementation specialists (@hermes, @aphrodite, @talos, @demeter, @prometheus), ALWAYS provide the exact, narrow test slice command for the changed file (e.g. `node --test tests/<slice>.test.mjs`, `pytest tests/test_<slice>.py -k <func>`).
|
|
71
|
+
- **Global Suites Forbidden in Dev Loops:** STRICTLY FORBID specialists from running global repository-wide test suites (`npm test`, `npm run test:all`, unconstrained `pytest`). Full suites cause massive timeouts, context bloating, and token waste. Full suites are reserved exclusively for the final release/PR gate.
|
|
72
|
+
- **Failure/Fix Loop (No Redundant Passes):** When fixing a failed test, specialists MUST rerun ONLY the failed test (`-k <func>` or `--test-name-pattern`). DO NOT repeat passing tests if code was unchanged.
|
|
73
|
+
- **Mapping Gaps & Scope Justification:** If no direct test file exists, report the mapping gap and run static lint/typecheck fallback (`npx biome check <file>` or `ruff check <file>`) plus nearest slice; never escalate silently to full suites. Any broader-than-slice run requires visible scope justification.
|
|
74
|
+
- **Observable Summary Receipt:** Require the specialist to return the structured receipt: `**tests:** [PASS|WARN|FAIL] <cmd> (targets: N, duration: <X>s, selection: <reason>)`.
|
|
75
|
+
|
|
63
76
|
## Fast-Path Direct Implementation (Default Coding Workflow)
|
|
64
77
|
|
|
65
78
|
- **Direct coding requests (features, bug fixes, refactoring, tests, hotfixes)**:
|
|
@@ -72,11 +85,19 @@ skills:
|
|
|
72
85
|
- **Planning via @athena is strictly gated**:
|
|
73
86
|
- ONLY invoke Athena when the user explicitly runs `/pantheon-deepwork` or explicitly requests a plan in the prompt ("plan this first", "create a plan", "planeje").
|
|
74
87
|
|
|
88
|
+
## Session Task Checklist & TUI Todo Panel (`todowrite`)
|
|
89
|
+
|
|
90
|
+
- **Live TUI Progress:** Whenever coordinating a workflow with two or more steps, or executing an architecture/deepwork plan:
|
|
91
|
+
- **Initialize:** Call `todowrite` immediately with the sequential task checklist (all items marked `pending`).
|
|
92
|
+
- **Transition:** As each phase or specialist delegation starts, update the active item to `in_progress` via `todowrite`.
|
|
93
|
+
- **Complete:** Upon verification or specialist return, mark the item `completed` (or `cancelled` if discarded) via `todowrite`.
|
|
94
|
+
- This ensures the active session's task status is visible in real-time in the Pantheon TUI Todo block.
|
|
95
|
+
|
|
75
96
|
## Golden Rule & Boundaries
|
|
76
97
|
|
|
77
|
-
**Coordinator and router.** Zeus never edits or implements code (delegates to @hermes, @aphrodite, @talos, etc.), but may perform bounded reads (≤50 lines) for fast routing without unnecessary subagent dispatches. Broad investigations or exploratory research go to @apollo.
|
|
98
|
+
**Coordinator and router.** Zeus never edits or implements code (delegates to @hermes, @aphrodite, @talos, etc.), but may perform bounded reads (≤50 lines) for fast routing without unnecessary subagent dispatches. Broad investigations or exploratory research go to @apollo. Routing and inspecting files across the `.pantheon/` directory (`.pantheon/**`, including plans, memory-bank, reports, configs) and active worktrees (`.pantheon-worktrees/**`) is fully permitted without restriction. Avoid broad recursive scans across `.git/**` or `node_modules/**`.
|
|
78
99
|
|
|
79
|
-
Zeus NEVER edits files (`edit: deny`, `hashline_edit: deny`). Zero
|
|
100
|
+
Zeus NEVER edits files (`edit: deny`, `hashline_edit: deny`). Zero terminal access (`bash: deny`, `shell: deny`). OpenCode's native terminal tool is `bash` (aliased to `shell`). Zeus has NO terminal access.
|
|
80
101
|
NEVER ask the user to type terminal commands — dispatch the specialist with bash permitted (@prometheus, @hermes, @talos).
|
|
81
102
|
|
|
82
103
|
## MCP call timing (when each integration fires)
|
|
@@ -11,7 +11,7 @@ All agents check each result against the request and evidence; remove unrelated
|
|
|
11
11
|
## subtask_summary
|
|
12
12
|
**files_changed:** [list of file paths, one per line]
|
|
13
13
|
**summary:** What was done, in 2-3 sentences
|
|
14
|
-
**tests:** [PASS
|
|
14
|
+
**tests:** [PASS | WARN | FAIL] `<command>` (targets: N, duration: <X>s, selection: <reason>)
|
|
15
15
|
**coverage:** X% (if applicable)
|
|
16
16
|
**tokens:** ~N input / ~M output (estimated)
|
|
17
17
|
**status:** complete | partial (reason) | escalated (reason)
|
|
@@ -1,13 +1,13 @@
|
|
|
1
1
|
---
|
|
2
|
-
description: "Test economy —
|
|
2
|
+
description: "Test economy — sub-10s test slicing; heavy suites are opt-in."
|
|
3
3
|
name: "Test Economy"
|
|
4
4
|
---
|
|
5
5
|
|
|
6
6
|
Write one focused test per behavior; do not duplicate coverage.
|
|
7
7
|
|
|
8
|
-
-
|
|
9
|
-
-
|
|
10
|
-
-
|
|
11
|
-
-
|
|
12
|
-
-
|
|
8
|
+
- **Test Slicing**: In dev loops, run ONLY the narrowest slice (<10s) for changed code (`node --test <slice>`, `pytest -k`); avoid unconstrained suites.
|
|
9
|
+
- **Repo Selector & Receipt**: Prefer repo selectors (`test:quick`, `test:slice`); record command, targets, duration, and selection reason in `subtask_summary`.
|
|
10
|
+
- **Failure/Fix Loops**: Rerun ONLY failed checks; never rerun passing tests without code changes.
|
|
11
|
+
- **Mapping Gaps**: If a file lacks tests, run static check fallback plus nearest slice.
|
|
12
|
+
- Slow install/E2E suites are opt-in, never per edit; reserve for release gates.
|
|
13
13
|
- Test economy does NOT weaken stronger gates: auth, security, data-integrity, and required coverage.
|
|
@@ -0,0 +1,303 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Auto-Permissions Engine — Intelligent Scope Governance.
|
|
3
|
+
*
|
|
4
|
+
* Intercepts OpenCode V2 permission evaluation events. When an action lands in
|
|
5
|
+
* `ask` status (permission prompt fatigue), it evaluates whether the requested
|
|
6
|
+
* tool or shell command is a safe, routine development action within the
|
|
7
|
+
* workspace.
|
|
8
|
+
*
|
|
9
|
+
* Rules:
|
|
10
|
+
* 1. Hard deny preservation: explicit host `deny` is NEVER overridden.
|
|
11
|
+
* 2. Catastrophic command rejection: dangerous operations (`rm -rf /`, `sudo`,
|
|
12
|
+
* arbitrary remote execution `curl | sh`, system path alteration) are strictly denied.
|
|
13
|
+
* 3. Read-only role isolation: read-only agents (`zeus`, `athena`, `apollo`, `gaia`)
|
|
14
|
+
* are never permitted to execute mutating tools or arbitrary shell commands.
|
|
15
|
+
* 4. Safe dev tool grant: routine developer commands (`node`, `python3`, `npm`,
|
|
16
|
+
* `git`, `mkdir`, `cat`, `ls`, local test scripts, etc.) and Pantheon native
|
|
17
|
+
* tools (`todowrite`, `pantheon_*`) are auto-approved without prompting.
|
|
18
|
+
* 5. Kill-switch: `PANTHEON_AUTO_PERMISSIONS=off` leaves `ask` untouched.
|
|
19
|
+
*
|
|
20
|
+
* @module auto-permissions
|
|
21
|
+
*/
|
|
22
|
+
|
|
23
|
+
export interface AutoPermissionEvent {
|
|
24
|
+
action?: unknown
|
|
25
|
+
resources?: unknown
|
|
26
|
+
agent?: unknown
|
|
27
|
+
sessionID?: unknown
|
|
28
|
+
effect?: unknown
|
|
29
|
+
message?: unknown
|
|
30
|
+
}
|
|
31
|
+
|
|
32
|
+
export interface AutoPermissionOptions {
|
|
33
|
+
cwd?: string
|
|
34
|
+
env?: Record<string, string | undefined>
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
export interface AutoPermissionResult {
|
|
38
|
+
decision:
|
|
39
|
+
| 'allow'
|
|
40
|
+
| 'deny'
|
|
41
|
+
| 'preserve-deny'
|
|
42
|
+
| 'preserve-allow'
|
|
43
|
+
| 'passthrough'
|
|
44
|
+
| 'disabled'
|
|
45
|
+
| 'unconstrained-suite'
|
|
46
|
+
reason?: string
|
|
47
|
+
}
|
|
48
|
+
|
|
49
|
+
/** Agents strictly forbidden from mutating workspace files or executing arbitrary shell commands. */
|
|
50
|
+
const READ_ONLY_AGENTS: ReadonlySet<string> = new Set(['zeus', 'athena', 'apollo', 'gaia'])
|
|
51
|
+
|
|
52
|
+
/** Allowed binary names for routine workspace development. */
|
|
53
|
+
const SAFE_DEV_BINARIES: ReadonlySet<string> = new Set([
|
|
54
|
+
'git',
|
|
55
|
+
'npm',
|
|
56
|
+
'npx',
|
|
57
|
+
'node',
|
|
58
|
+
'python',
|
|
59
|
+
'python3',
|
|
60
|
+
'pytest',
|
|
61
|
+
'ruff',
|
|
62
|
+
'cat',
|
|
63
|
+
'ls',
|
|
64
|
+
'mkdir',
|
|
65
|
+
'cp',
|
|
66
|
+
'mv',
|
|
67
|
+
'touch',
|
|
68
|
+
'grep',
|
|
69
|
+
'find',
|
|
70
|
+
'head',
|
|
71
|
+
'tail',
|
|
72
|
+
'diff',
|
|
73
|
+
'echo',
|
|
74
|
+
'which',
|
|
75
|
+
'tar',
|
|
76
|
+
'zip',
|
|
77
|
+
'unzip',
|
|
78
|
+
'curl',
|
|
79
|
+
'jq',
|
|
80
|
+
'sed',
|
|
81
|
+
'awk',
|
|
82
|
+
'biome',
|
|
83
|
+
'tsc',
|
|
84
|
+
'pnpm',
|
|
85
|
+
'yarn',
|
|
86
|
+
'bun',
|
|
87
|
+
'cargo',
|
|
88
|
+
'go',
|
|
89
|
+
'sh',
|
|
90
|
+
'bash',
|
|
91
|
+
])
|
|
92
|
+
|
|
93
|
+
/** Check if auto-permissions engine is enabled in environment. */
|
|
94
|
+
export function autoPermissionsEnabled(
|
|
95
|
+
env: Record<string, string | undefined> = process.env,
|
|
96
|
+
): boolean {
|
|
97
|
+
const val = (env.PANTHEON_AUTO_PERMISSIONS ?? '').trim().toLowerCase()
|
|
98
|
+
return val !== 'off' && val !== '0' && val !== 'false'
|
|
99
|
+
}
|
|
100
|
+
|
|
101
|
+
/** Check if a command matches high-risk destructive or privilege escalation patterns. */
|
|
102
|
+
export function isDangerousCommand(command: string): boolean {
|
|
103
|
+
const trimmed = command.trim()
|
|
104
|
+
if (!trimmed) return false
|
|
105
|
+
|
|
106
|
+
// Privilege escalation
|
|
107
|
+
if (/\bsudo\b/.test(trimmed)) return true
|
|
108
|
+
|
|
109
|
+
// Destructive recursive deletion targeting root, all root items, or user home
|
|
110
|
+
if (
|
|
111
|
+
/\brm\s+(-[a-zA-Z]*r[a-zA-Z]*f|-[a-zA-Z]*f[a-zA-Z]*r|--force)\s+(\/\s*$|\/\*\s*$|~\/?\s*$|\$HOME)/.test(
|
|
112
|
+
trimmed,
|
|
113
|
+
)
|
|
114
|
+
)
|
|
115
|
+
return true
|
|
116
|
+
|
|
117
|
+
// Remote execution via piping to shell
|
|
118
|
+
if (/(curl|wget)\b.*\|\s*(ba)?sh\b/.test(trimmed)) return true
|
|
119
|
+
|
|
120
|
+
// Raw disk formatting or writing
|
|
121
|
+
if (/\bmkfs(\.[a-z0-9]+)?\b/.test(trimmed)) return true
|
|
122
|
+
if (/\bdd\s+if=/.test(trimmed)) return true
|
|
123
|
+
|
|
124
|
+
// Global permissions tampering
|
|
125
|
+
if (/\bchmod\s+(-[a-zA-Z]*R[a-zA-Z]*\s+)?777\s+\//.test(trimmed)) return true
|
|
126
|
+
|
|
127
|
+
// Sensitive system directory writes
|
|
128
|
+
if (/\b(>\s*|\bcp\b.*|\bmv\b.*)\s+\/(etc|boot|sys|proc|root|usr)(\/|$)/.test(trimmed)) return true
|
|
129
|
+
|
|
130
|
+
return false
|
|
131
|
+
}
|
|
132
|
+
|
|
133
|
+
/** Check if a command is a standard dev command within workspace. */
|
|
134
|
+
export function isSafeDevCommand(command: string): boolean {
|
|
135
|
+
const trimmed = command.trim()
|
|
136
|
+
if (!trimmed) return false
|
|
137
|
+
if (isDangerousCommand(trimmed)) return false
|
|
138
|
+
|
|
139
|
+
// Extract the primary command tokens (before flags/arguments, handling simple env prefixes)
|
|
140
|
+
const segments = trimmed.split(/[|&;]/)
|
|
141
|
+
for (const rawSeg of segments) {
|
|
142
|
+
const seg = rawSeg.trim()
|
|
143
|
+
if (!seg) continue
|
|
144
|
+
|
|
145
|
+
const tokens = seg.split(/\s+/).filter(Boolean)
|
|
146
|
+
let binary = tokens[0] ?? ''
|
|
147
|
+
|
|
148
|
+
// Skip environment assignments like FOO=bar node ...
|
|
149
|
+
let tokenIdx = 0
|
|
150
|
+
while (binary.includes('=') && tokenIdx < tokens.length - 1) {
|
|
151
|
+
tokenIdx += 1
|
|
152
|
+
binary = tokens[tokenIdx] ?? ''
|
|
153
|
+
}
|
|
154
|
+
|
|
155
|
+
const cleanBinary = binary.replace(/^.*\//, '')
|
|
156
|
+
if (!SAFE_DEV_BINARIES.has(cleanBinary)) {
|
|
157
|
+
return false
|
|
158
|
+
}
|
|
159
|
+
}
|
|
160
|
+
|
|
161
|
+
return true
|
|
162
|
+
}
|
|
163
|
+
|
|
164
|
+
/**
|
|
165
|
+
* Detect unconstrained repository-wide test suite invocations.
|
|
166
|
+
*
|
|
167
|
+
* During active dev/fix/implementation loops, unconstrained test suites cause
|
|
168
|
+
* massive latency, timeouts, and token waste (Issue #437). Auto-permissions
|
|
169
|
+
* holds these in 'ask' so agents cannot silently burn minutes running hundreds
|
|
170
|
+
* of unrelated tests without explicit user intent or visible scope justification.
|
|
171
|
+
*/
|
|
172
|
+
export function isUnconstrainedTestSuite(command: string): boolean {
|
|
173
|
+
const trimmed = command.trim()
|
|
174
|
+
if (!trimmed) return false
|
|
175
|
+
|
|
176
|
+
// Unconstrained npm test (without arguments or slice flags)
|
|
177
|
+
// Matches: npm test, npm run test, npm run test:all, pnpm test, yarn test, bun test
|
|
178
|
+
if (/^(npm|pnpm|yarn|bun)\s+(run\s+)?(test|test:all)(\s+--)?\s*$/.test(trimmed)) {
|
|
179
|
+
return true
|
|
180
|
+
}
|
|
181
|
+
|
|
182
|
+
// Unconstrained pytest (without -k, -m, or specific test path)
|
|
183
|
+
if (/^(python3?\s+-m\s+)?pytest\s*$/.test(trimmed)) {
|
|
184
|
+
return true
|
|
185
|
+
}
|
|
186
|
+
|
|
187
|
+
// Unconstrained vitest or jest (without file path or filter)
|
|
188
|
+
if (/^(npx\s+)?(vitest|jest)(\s+run)?\s*$/.test(trimmed)) {
|
|
189
|
+
return true
|
|
190
|
+
}
|
|
191
|
+
|
|
192
|
+
// Unconstrained cargo test
|
|
193
|
+
if (/^cargo\s+test\s*$/.test(trimmed)) {
|
|
194
|
+
return true
|
|
195
|
+
}
|
|
196
|
+
|
|
197
|
+
return false
|
|
198
|
+
}
|
|
199
|
+
|
|
200
|
+
/**
|
|
201
|
+
* Evaluate an OpenCode V2 permission evaluation event.
|
|
202
|
+
* If safe to auto-allow, sets `event.effect = 'allow'`.
|
|
203
|
+
*/
|
|
204
|
+
export function evaluateAutoPermission(
|
|
205
|
+
event: AutoPermissionEvent,
|
|
206
|
+
options: AutoPermissionOptions = {},
|
|
207
|
+
): AutoPermissionResult {
|
|
208
|
+
const env = options.env ?? process.env
|
|
209
|
+
if (!autoPermissionsEnabled(env)) {
|
|
210
|
+
return { decision: 'disabled' }
|
|
211
|
+
}
|
|
212
|
+
|
|
213
|
+
if (event.effect === 'deny') {
|
|
214
|
+
return { decision: 'preserve-deny' }
|
|
215
|
+
}
|
|
216
|
+
|
|
217
|
+
if (event.effect === 'allow') {
|
|
218
|
+
return { decision: 'preserve-allow' }
|
|
219
|
+
}
|
|
220
|
+
|
|
221
|
+
const action = String(event.action ?? '')
|
|
222
|
+
.toLowerCase()
|
|
223
|
+
.trim()
|
|
224
|
+
const agent = String(event.agent ?? '')
|
|
225
|
+
.toLowerCase()
|
|
226
|
+
.trim()
|
|
227
|
+
|
|
228
|
+
// 1. Native Pantheon tools and Todo tools are always auto-allowed
|
|
229
|
+
if (
|
|
230
|
+
action === 'todowrite' ||
|
|
231
|
+
action === 'todoread' ||
|
|
232
|
+
action === 'pantheon_deepwork' ||
|
|
233
|
+
action.startsWith('pantheon_')
|
|
234
|
+
) {
|
|
235
|
+
event.effect = 'allow'
|
|
236
|
+
return { decision: 'allow', reason: 'pantheon-native-tool' }
|
|
237
|
+
}
|
|
238
|
+
|
|
239
|
+
// 2. Read-only agents are strictly prohibited from mutating tools or shell
|
|
240
|
+
if (READ_ONLY_AGENTS.has(agent)) {
|
|
241
|
+
if (['edit', 'write', 'hashline_edit', 'bash', 'shell'].includes(action)) {
|
|
242
|
+
event.effect = 'deny'
|
|
243
|
+
event.message = `Pantheon auto-permissions denied: read-only agent '${agent}' cannot execute mutating action '${action}'`
|
|
244
|
+
return { decision: 'deny', reason: 'read-only-role-violation' }
|
|
245
|
+
}
|
|
246
|
+
}
|
|
247
|
+
|
|
248
|
+
// 3. Read-only inspection tools are safe within workspace (excluding network/external URLs)
|
|
249
|
+
if (['read', 'glob', 'grep', 'webfetch', 'question'].includes(action)) {
|
|
250
|
+
const firstResource =
|
|
251
|
+
Array.isArray(event.resources) && typeof event.resources[0] === 'string'
|
|
252
|
+
? event.resources[0]
|
|
253
|
+
: ''
|
|
254
|
+
if (firstResource.startsWith('http://') || firstResource.startsWith('https://')) {
|
|
255
|
+
return { decision: 'passthrough' }
|
|
256
|
+
}
|
|
257
|
+
event.effect = 'allow'
|
|
258
|
+
return { decision: 'allow', reason: 'safe-read-action' }
|
|
259
|
+
}
|
|
260
|
+
|
|
261
|
+
// 4. Shell commands
|
|
262
|
+
if (action === 'bash' || action === 'shell') {
|
|
263
|
+
const rawCmd =
|
|
264
|
+
Array.isArray(event.resources) && typeof event.resources[0] === 'string'
|
|
265
|
+
? event.resources[0]
|
|
266
|
+
: ''
|
|
267
|
+
|
|
268
|
+
if (rawCmd) {
|
|
269
|
+
if (isDangerousCommand(rawCmd)) {
|
|
270
|
+
event.effect = 'deny'
|
|
271
|
+
event.message =
|
|
272
|
+
'Pantheon auto-permissions denied: destructive or privilege-escalating command'
|
|
273
|
+
return { decision: 'deny', reason: 'dangerous-command' }
|
|
274
|
+
}
|
|
275
|
+
|
|
276
|
+
if (isUnconstrainedTestSuite(rawCmd)) {
|
|
277
|
+
event.effect = 'ask'
|
|
278
|
+
event.message =
|
|
279
|
+
'Pantheon test-economy: unconstrained test suite detected in dev loop. ' +
|
|
280
|
+
'Use a focused test slice (sub-10s) or provide scope justification.'
|
|
281
|
+
return {
|
|
282
|
+
decision: 'unconstrained-suite',
|
|
283
|
+
reason: 'unconstrained-test-suite-forbidden-in-dev-loop',
|
|
284
|
+
}
|
|
285
|
+
}
|
|
286
|
+
|
|
287
|
+
if (isSafeDevCommand(rawCmd)) {
|
|
288
|
+
event.effect = 'allow'
|
|
289
|
+
return { decision: 'allow', reason: 'safe-dev-command' }
|
|
290
|
+
}
|
|
291
|
+
}
|
|
292
|
+
}
|
|
293
|
+
|
|
294
|
+
// 5. File mutations for implementation agents
|
|
295
|
+
if (['edit', 'write', 'hashline_edit'].includes(action)) {
|
|
296
|
+
if (!READ_ONLY_AGENTS.has(agent)) {
|
|
297
|
+
event.effect = 'allow'
|
|
298
|
+
return { decision: 'allow', reason: 'implementer-workspace-edit' }
|
|
299
|
+
}
|
|
300
|
+
}
|
|
301
|
+
|
|
302
|
+
return { decision: 'passthrough' }
|
|
303
|
+
}
|
|
@@ -248,6 +248,17 @@ export function createDeepworkTool(cwd = process.cwd()): V2ToolDef {
|
|
|
248
248
|
|
|
249
249
|
const projection = generateChatProjection(initialPlan, effectiveCwd)
|
|
250
250
|
const approvalQ = buildApprovalQuestion(initialPlan)
|
|
251
|
+
const todoPayload = {
|
|
252
|
+
todos: tasks.map((t) => ({
|
|
253
|
+
content: `[${t.id}] ${t.title}`,
|
|
254
|
+
status:
|
|
255
|
+
t.state === 'done'
|
|
256
|
+
? 'completed'
|
|
257
|
+
: t.state === 'in_progress'
|
|
258
|
+
? 'in_progress'
|
|
259
|
+
: 'pending',
|
|
260
|
+
})),
|
|
261
|
+
}
|
|
251
262
|
|
|
252
263
|
return {
|
|
253
264
|
output: [
|
|
@@ -257,6 +268,9 @@ export function createDeepworkTool(cwd = process.cwd()): V2ToolDef {
|
|
|
257
268
|
'',
|
|
258
269
|
projection,
|
|
259
270
|
'',
|
|
271
|
+
'**TUI Todo Sync:** Espelhe as tarefas chamando `todowrite` com:',
|
|
272
|
+
`\`\`\`json\n${JSON.stringify(todoPayload, null, 2)}\n\`\`\``,
|
|
273
|
+
'',
|
|
260
274
|
'**Pergunta de Aprovação Nativa (Gate 1):**',
|
|
261
275
|
`\`${JSON.stringify(approvalQ)}\``,
|
|
262
276
|
].join('\n'),
|