pantheon-opencode 2.0.0-beta.6 → 2.0.0
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/AGENTS.global.md +7 -2
- package/AGENTS.md +8 -8
- package/CHANGELOG.md +99 -0
- package/README.md +171 -510
- package/README.pt-BR.md +179 -149
- package/bin/pantheon-init.mjs +14 -1
- package/commands/pantheon-audit.md +1 -1
- package/commands/pantheon-deepwork.md +7 -2
- package/commands/pantheon-optimize.md +41 -63
- package/commands/pantheon.md +10 -5
- package/package.json +4 -3
- package/plugin.json +2 -2
- package/prompts/optimize-database.prompt.md +5 -5
- package/prompts/orchestrate-with-zeus.prompt.md +82 -290
- package/prompts/plan-architecture.prompt.md +6 -6
- package/prompts/subtask.prompt.md +1 -1
- package/pyproject.toml +1 -1
- package/scripts/build-agents-md.mjs +10 -10
- package/scripts/build-plugin.mjs +26 -0
- package/scripts/doctor.mjs +27 -11
- package/scripts/hooks/format-multi-language.sh +30 -48
- package/scripts/install/opencode-version.mjs +17 -5
- package/scripts/install/opencode.mjs +8 -1
- package/scripts/jev-live-probe.mjs +1 -1
- package/scripts/optimize.mjs +322 -0
- package/scripts/prune.mjs +105 -14
- package/scripts/release-gates.mjs +19 -5
- package/scripts/skill-search.mjs +27 -3
- package/scripts/test-opencode-sandbox.sh +44 -15
- package/scripts/update.mjs +29 -17
- package/skills-lock.json +15 -13
- package/src/agents/aphrodite.md +17 -14
- package/src/agents/apollo.md +10 -207
- package/src/agents/athena.md +30 -34
- package/src/agents/demeter.md +2 -0
- package/src/agents/hephaestus.md +1 -0
- package/src/agents/hermes.md +21 -95
- package/src/agents/iris.md +3 -0
- package/src/agents/mnemosyne.md +1 -1
- package/src/agents/nyx.md +2 -1
- package/src/agents/prometheus.md +2 -203
- package/src/agents/talos.md +5 -0
- package/src/agents/themis.md +93 -84
- package/src/agents/zeus.md +44 -263
- package/src/instructions/agent-return-format.instructions.md +1 -1
- package/src/instructions/memory-protocol.instructions.md +5 -1
- package/src/instructions/test-economy.instructions.md +1 -0
- package/src/instructions/zeus-anti-stall.instructions.md +3 -3
- package/src/instructions/zeus-communication-rules.instructions.md +1 -1
- package/src/instructions/zeus-council-synthesis.instructions.md +15 -3
- package/src/instructions/zeus-timeout-retry.instructions.md +10 -10
- package/src/pantheon/model-command.ts +12 -12
- package/src/pantheon/pantheon-decision-catalog.ts +8 -0
- package/src/pantheon/skill-search-tool.ts +193 -0
- package/src/pantheon/tools.ts +6 -0
- package/src/pantheon/tui-version-check.ts +1 -1
- package/src/pantheon/update-notifier.ts +13 -8
- package/src/pantheon/vision.ts +20 -20
- package/src/plugin/index.js +14594 -0
- package/src/plugins/tui/dist/tui.d.ts +15 -5
- package/src/plugins/tui/dist/tui.d.ts.map +1 -1
- package/src/plugins/tui/dist/tui.js +1761 -556
- package/src/plugins/tui/dist/tui.js.map +1 -1
- package/src/plugins/tui/package-lock.json +2 -2
- package/src/plugins/tui/package.json +1 -1
- package/src/plugins/tui/src/autonomy-state.ts +28 -16
- package/src/plugins/tui/src/context-gauge.ts +64 -5
- package/src/plugins/tui/src/decision-router-state.ts +49 -2
- package/src/plugins/tui/src/index.tsx +69 -26
- package/src/plugins/tui/src/pantheon-settings-state.ts +99 -0
- package/src/plugins/tui/src/pantheon-settings.tsx +771 -125
- package/src/plugins/tui/src/tui-prefs.ts +266 -12
- package/src/plugins/tui/src/update-notice.ts +9 -6
- package/src/plugins/tui/src/update-status.ts +27 -8
- package/src/plugins/tui/src/whats-new-open.ts +4 -4
- package/src/plugins/tui/src/whats-new.tsx +119 -17
- package/src/routing.yml +9 -0
- package/src/skills/README.md +35 -25
- package/src/skills/agent-coordination/SKILL.md +6 -6
- package/src/skills/artifact-management/SKILL.md +43 -39
- package/src/skills/artifact-management/references/artifact-templates.md +103 -48
- package/src/skills/auto-continue/references/auto-continue-details.md +13 -13
- package/src/skills/code-review-checklist/SKILL.md +6 -6
- package/src/skills/fast-path-execution/SKILL.md +54 -0
- package/src/skills/incremental-implementation/SKILL.md +13 -13
- package/src/skills/memory-bank/references/anti-patterns.md +3 -3
- package/src/skills/prompt-refinery/SKILL.md +51 -65
- package/src/skills/spec-driven-development/SKILL.md +12 -14
- package/src/skills/tdd-with-agents/SKILL.md +23 -10
- package/src/skills/test-slicing/SKILL.md +56 -0
- package/src/skills/verification-planning/SKILL.md +4 -3
- package/src/skills/visual-review-pipeline/SKILL.md +1 -1
- package/src/skills/visual-review-pipeline/references/workflow-steps.md +1 -1
package/README.md
CHANGED
|
@@ -1,9 +1,9 @@
|
|
|
1
1
|
# Pantheon
|
|
2
2
|
|
|
3
|
-
**A clearer way to work with OpenCode on real software projects.** Pantheon
|
|
4
|
-
brings planning, implementation, review, and documentation into one guided
|
|
5
|
-
experience. It is an OpenCode plugin and installer for developers
|
|
6
|
-
want useful structure without giving up control of their code.
|
|
3
|
+
**A clearer, faster way to work with OpenCode on real software projects.** Pantheon
|
|
4
|
+
brings planning, implementation, review, and documentation into one guided,
|
|
5
|
+
token-efficient experience. It is an OpenCode plugin and installer for developers
|
|
6
|
+
and teams who want useful structure without giving up speed or control of their code.
|
|
7
7
|
|
|
8
8
|
[Português (Brasil)](README.pt-BR.md) ·
|
|
9
9
|
[Repository](https://github.com/ils15/pantheon-opencode) · [MIT License](LICENSE)
|
|
@@ -15,552 +15,213 @@ want useful structure without giving up control of their code.
|
|
|
15
15
|
## What is it?
|
|
16
16
|
|
|
17
17
|
Pantheon is a companion for [OpenCode](https://opencode.ai/) that helps you move
|
|
18
|
-
from an idea to a reviewed change. Fourteen specialized agents share one
|
|
19
|
-
workflow: plan, implement, check, review, and record what mattered.
|
|
20
|
-
charge of every decision and every merge.
|
|
18
|
+
from an idea to a reviewed, verified change. Fourteen specialized agents share one
|
|
19
|
+
streamlined workflow: plan, implement, check, review, and record what mattered.
|
|
20
|
+
You stay in charge of every decision and every merge.
|
|
21
21
|
|
|
22
22
|
```mermaid
|
|
23
23
|
flowchart LR
|
|
24
|
-
U["
|
|
24
|
+
U["You"] --> Z["Zeus<br/>orchestrator"]
|
|
25
25
|
Z -->|"delegates"| S["14 specialists<br/>hermes · aphrodite · demeter · themis …"]
|
|
26
|
-
Z -->|"ambiguous route"| J["
|
|
26
|
+
Z -->|"ambiguous route"| J["Jev decision router<br/>typed · budgeted · reversible"]
|
|
27
27
|
J -->|"orders · ranks · verifies"| S
|
|
28
|
-
S -->|"last resort"| M["
|
|
29
|
-
Z --> G["
|
|
30
|
-
G -->|"high risk"| H["
|
|
28
|
+
S -->|"last resort"| M["Native MCP<br/>memory · persistence · vision<br/>code-mode · resources"]
|
|
29
|
+
Z --> G["J1 risk gate<br/>off / shadow / warn / enforce"]
|
|
30
|
+
G -->|"high risk"| H["Human approval"]
|
|
31
31
|
```
|
|
32
32
|
|
|
33
|
-
**The escalation ladder every agent follows
|
|
34
|
-
agent files): session context first → the free local selector → one typed Jev
|
|
35
|
-
question when genuinely ambiguous → a native MCP call only as the LAST resort
|
|
36
|
-
for state the session cannot reconstruct.
|
|
33
|
+
**The escalation ladder every agent follows**: session context first → free local selector → one typed Jev decision when genuinely ambiguous → native MCP call only as the last resort for state the session cannot reconstruct.
|
|
37
34
|
|
|
38
35
|
```mermaid
|
|
39
36
|
flowchart LR
|
|
40
|
-
R["
|
|
41
|
-
T0 -->|yes| X["
|
|
42
|
-
T0 -->|"vague /<br/>multi-concern"| T1["
|
|
37
|
+
R["rough request"] --> T0{"clear &<br/>bounded?"}
|
|
38
|
+
T0 -->|yes| X["execute directly"]
|
|
39
|
+
T0 -->|"vague /<br/>multi-concern"| T1["Tier 1 — internal<br/>minimal adjust ≤15 lines<br/>(silent, no approval)"]
|
|
43
40
|
T1 --> T2q{">> given?"}
|
|
44
|
-
T2q -->|yes| T2["
|
|
41
|
+
T2q -->|yes| T2["Tier 2 — full contract<br/>6 sections, confirm-or-adjust"]
|
|
45
42
|
T2q -->|no| X
|
|
46
43
|
```
|
|
47
44
|
|
|
48
|
-
**Skill discovery runs a trust cascade** — canonical first, popularity never
|
|
49
|
-
compensates for provenance:
|
|
45
|
+
**Skill discovery runs a trust cascade** — canonical first, popularity never compensates for provenance:
|
|
50
46
|
|
|
51
47
|
```mermaid
|
|
52
48
|
flowchart LR
|
|
53
|
-
L["
|
|
54
|
-
C1 -->|"no fit"| C2["
|
|
55
|
-
C2 -->|"no fit"| C3["
|
|
56
|
-
C3 --> J["
|
|
57
|
-
J --> P["
|
|
49
|
+
L["installed pool<br/>(always first)"] -->|"gap"| C1["anthropics/skills<br/>canonical"]
|
|
50
|
+
C1 -->|"no fit"| C2["skills.sh<br/>install-count signal"]
|
|
51
|
+
C2 -->|"no fit"| C3["GitHub raw<br/>widest, weakest"]
|
|
52
|
+
C3 --> J["Jev orders<br/>the candidates"]
|
|
53
|
+
J --> P["YOU review<br/>and run the install"]
|
|
58
54
|
```
|
|
59
55
|
|
|
60
56
|
## Why use it?
|
|
61
57
|
|
|
62
|
-
- **
|
|
63
|
-
- **
|
|
64
|
-
|
|
65
|
-
- **
|
|
66
|
-
|
|
67
|
-
- **You stay in charge:** Pantheon supports your judgment; it does not replace
|
|
68
|
-
it.
|
|
69
|
-
- **Right-sized coordination:** bounded work goes to one specialist, who
|
|
70
|
-
gathers the needed context. Planning, separate discovery, parallel work, and
|
|
71
|
-
review are added when scope and risk call for them. `/pantheon-deepwork`
|
|
72
|
-
remains the checkpointed workflow for heavy multi-phase work.
|
|
58
|
+
- **Fast-Track execution:** Bounded micro-edits (<= 2 files, <= 50 lines) bypass all heavy ceremony and verify in < 10 seconds.
|
|
59
|
+
- **Mandatory test slicing:** Sub-second targeted test execution instead of running sluggish repository-wide test suites during dev loops.
|
|
60
|
+
- **Assertive council (`/pantheon`):** Parallel specialist consultations resolved decisively via domain authority and JEV multi-criteria decisions without endless debate turns.
|
|
61
|
+
- **You stay in charge:** Pantheon supports your judgment; it never silently alters permissions or runs unapproved destructive actions.
|
|
62
|
+
- **Non-destructive upgrades:** Updating Pantheon (`pantheon-opencode update`) refreshes skills, MCPs, and instructions while strictly preserving all user models, presets, permissions, and custom configs.
|
|
73
63
|
|
|
74
64
|
## Start in 2 minutes
|
|
75
65
|
|
|
76
|
-
Requirements: [OpenCode](https://opencode.ai/docs/) **2.x** and Node.js
|
|
77
|
-
**>=26.4.0** for the CLI and tooling. **OpenCode 2.0.22 is known broken**; see
|
|
78
|
-
[Compatibility](docs/COMPATIBILITY.md).
|
|
66
|
+
Requirements: [OpenCode](https://opencode.ai/docs/) **2.x** and Node.js **>=26.4.0**. (See [Compatibility](docs/COMPATIBILITY.md)).
|
|
79
67
|
|
|
80
|
-
|
|
68
|
+
### 1. Install
|
|
81
69
|
|
|
70
|
+
**Option A — Quick Plugin Add (Global or Project):**
|
|
82
71
|
```bash
|
|
83
72
|
opencode plugin add pantheon-opencode
|
|
84
|
-
#
|
|
85
|
-
opencode
|
|
73
|
+
# Or pin a specific version:
|
|
74
|
+
opencode plugin add pantheon-opencode@2.0.0-rc.1
|
|
86
75
|
```
|
|
87
76
|
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
`prompts` are opt-in via `--components`. Since beta.5 the TTY run is
|
|
94
|
-
prompt-light: banner → per-file plan preview → ONE final authorization
|
|
95
|
-
(default **no**) — no component/target/decision-model questions. Everything
|
|
96
|
-
those prompts used to configure lives in `/pantheon-settings` (gates,
|
|
97
|
-
decision router, model preset, TUI personalization) or in explicit flags
|
|
98
|
-
(`--components`, `--project`, `--decision-model`,
|
|
99
|
-
`--decision-monthly-usd`, `--decision-per-decision-usd`). Without a TTY,
|
|
100
|
-
`init` is preview-only unless `--yes` explicitly authorizes the apply;
|
|
101
|
-
`--yes` skips prompts but still prints the preview and checks for drift.
|
|
102
|
-
|
|
103
|
-
`npm run setup` is a separate repository setup wrapper. On a detected
|
|
104
|
-
OpenCode 2.x host it reviews the global config before installer init: legacy
|
|
105
|
-
permission rules can be migrated after explicit review. Config migration and
|
|
106
|
-
installer changes are preflighted before one final authorization; `--yes`
|
|
107
|
-
alone does not authorize setup writes, and headless/CI setup remains
|
|
108
|
-
preview-only. Direct `npx pantheon-opencode init` does not run that
|
|
109
|
-
config-migration review. See the [installation guide](docs/INSTALLATION.md)
|
|
110
|
-
for layer handling, conflict choices, backups, and recovery.
|
|
111
|
-
Existing native `permissions` arrays remain authoritative. Unsafe MCP
|
|
112
|
-
wildcard policies and ineffective `ask` rules require explicit resolution;
|
|
113
|
-
third-party flat MCP entries in `opencode.jsonc` must be converted to
|
|
114
|
-
`mcp.servers` or removed before setup.
|
|
115
|
-
|
|
116
|
-
Global `init` and `setup` use `$XDG_CONFIG_HOME/opencode` when
|
|
117
|
-
`XDG_CONFIG_HOME` is absolute; otherwise they use `~/.config/opencode`.
|
|
118
|
-
Relative `XDG_CONFIG_HOME` values are rejected before writes.
|
|
119
|
-
|
|
120
|
-
Then describe the outcome you want:
|
|
77
|
+
**Option B — Global CLI Install with Full MCP Tooling:**
|
|
78
|
+
```bash
|
|
79
|
+
npm install -g pantheon-opencode@beta
|
|
80
|
+
pantheon-opencode init --yes
|
|
81
|
+
```
|
|
121
82
|
|
|
122
|
-
|
|
123
|
-
|
|
83
|
+
### 2. Verify Health
|
|
84
|
+
```bash
|
|
85
|
+
pantheon-opencode doctor
|
|
86
|
+
```
|
|
87
|
+
Runs a 28-point diagnostic checking MCP initialization, Python virtual environments, Node version, permissions, and config integrity.
|
|
88
|
+
|
|
89
|
+
### 3. Updating Pantheon Without Overwriting Settings
|
|
90
|
+
When a new version is released:
|
|
91
|
+
```bash
|
|
92
|
+
pantheon-opencode update
|
|
93
|
+
```
|
|
94
|
+
The updater verifies package identity, pulls the latest release, and runs a non-destructive configuration refresh. **Your custom models, active preset, custom permissions, plugin options, and user-defined AGENTS.md sections are preserved verbatim.**
|
|
95
|
+
|
|
96
|
+
---
|
|
97
|
+
|
|
98
|
+
## Optimal Workflow & Execution Modes
|
|
99
|
+
|
|
100
|
+
Pantheon balances speed and rigor through three distinct execution modes:
|
|
101
|
+
|
|
102
|
+
```mermaid
|
|
103
|
+
flowchart TD
|
|
104
|
+
Req["User Request"] --> Size{"Scope / Risk"}
|
|
105
|
+
Size -->|"≤ 2 files · ≤ 50 lines<br/>routine fix / typo / CSS"| Fast["Fast-Track<br/>(Sub-10s · Inline Test Slicing)"]
|
|
106
|
+
Size -->|"Architecture / Tech choice<br/>Security / Multi-stakeholder"| Council["Council Synthesis (/pantheon)<br/>(2-3 Specialists in Parallel · Domain Authority)"]
|
|
107
|
+
Size -->|"Multi-phase feature<br/>Heavy refactor · Breaking changes"| Deep["Deepwork (/pantheon-deepwork)<br/>(Athena Plan · Phase Gates · Checkpoints)"]
|
|
124
108
|
```
|
|
125
109
|
|
|
126
|
-
|
|
127
|
-
|
|
128
|
-
|
|
129
|
-
|
|
130
|
-
|
|
131
|
-
|
|
132
|
-
|
|
133
|
-
|
|
134
|
-
|
|
135
|
-
|
|
136
|
-
see [Quick Start](docs/QUICKSTART.md).
|
|
137
|
-
|
|
138
|
-
## What's included?
|
|
139
|
-
|
|
140
|
-
- **14 specialized agents** with one orchestrator, implementation specialists,
|
|
141
|
-
a quality gate, and supporting roles.
|
|
142
|
-
- **23 skills** and reusable instructions for planning, building, reviewing,
|
|
143
|
-
and documenting work.
|
|
144
|
-
- **5 local MCP servers** for project memory, persistence, resources, confined
|
|
145
|
-
script execution, and vision.
|
|
146
|
-
- **A selective, advisory decision router** with structured local results and
|
|
147
|
-
deterministic per-turn skill/specialist recommendations when enabled.
|
|
148
|
-
Pantheon-owned direct-provider egress is inactive; native egress exists only
|
|
149
|
-
via the explicit, Zeus-only `pantheon_jev_decide` advisory tool, which calls
|
|
150
|
-
the host's native Decisions API (TypeSafe Jev via OpenCode Console or a
|
|
151
|
-
configured decision model) behind the user's consent and budget gates — it
|
|
152
|
-
never dispatches agents, runs skills, or changes permissions. This does not
|
|
153
|
-
change the separately documented Vision integration. See the
|
|
154
|
-
[router reference](docs/reference/local-decision-router.md) and the
|
|
155
|
-
[Native Decisions (Jev)](docs/reference/native-decisions.md).
|
|
156
|
-
- **An optional TUI** with a dispatch sidebar block and a full-width
|
|
157
|
-
`/dispatch` panel. The sidebar keeps the Pantheon version, the active
|
|
158
|
-
preset/source, git state, and a `Context` block (tokens, a percent gauge,
|
|
159
|
-
`$<spent> spent`, and a muted `cache <n>% · <n.n> tok/s` line); the native
|
|
160
|
-
prompt footer keeps its own line, and Pantheon only appends
|
|
161
|
-
`▸<preset> · ✦<n> ativos` while delegations are running. When the Pantheon
|
|
162
|
-
version changes since the last TUI start, a one-time dismissable
|
|
163
|
-
"What's new" dialog lists that version's highlights from the shipped
|
|
164
|
-
changelog — once per version, `esc`/`q` dismisses; see
|
|
165
|
-
[What's new window](docs/reference/whats-new.md). `/pantheon-settings`
|
|
166
|
-
(also `Pantheon: Settings` in the command palette) opens the sectioned
|
|
167
|
-
settings window — `tab` cycles sections, `up`/`down` selects,
|
|
168
|
-
`enter` applies, `esc`/`q` closes — and since beta.4 it is a FULL-SCREEN
|
|
169
|
-
tabbed window (the OpenCode Quota shape): `←/→` cycles a row's value into a
|
|
170
|
-
staged draft, `ctrl+s` saves every staged change, `ctrl+r` resets the tab,
|
|
171
|
-
and a warning line surfaces every refusal:
|
|
172
|
-
|
|
173
|
-
- **Gates & Overrides** — toggle the named behaviors, writing the global
|
|
174
|
-
`pantheon-overrides.json` the installer and doctor consume. A corrupt
|
|
175
|
-
overrides file refuses writes (repair with `doctor --fix`); entries in the
|
|
176
|
-
project `.pantheon/overrides.json` win at runtime and are shown read-only.
|
|
177
|
-
- **Decision Model** — enable/disable the advisory decision router live
|
|
178
|
-
(the runtime re-reads the config per call). Egress consent stays
|
|
179
|
-
CLI-owned (`decision-router setup-model`).
|
|
180
|
-
- **Model Preset** — activate any routing.yml preset globally (atomic
|
|
181
|
-
write with a `.bak` backup); `none` clears the marker. A project
|
|
182
|
-
`.pantheon/active-preset.json` wins and shows read-only.
|
|
183
|
-
- **J1 Risk Gate** — ladder mode/threshold status; env-only config
|
|
184
|
-
(`PANTHEON_J1_*`), read at startup, never written here.
|
|
185
|
-
- **Updates** — channel and a `Check now` registry probe for the running
|
|
186
|
-
channel; adopt with `pantheon-opencode update`.
|
|
187
|
-
- **Doctor** — file-derived health: install-version drift, Node floor,
|
|
188
|
-
overrides and decision-config validation. The full check (MCP smoke,
|
|
189
|
-
venv pins, git) remains `pantheon-opencode doctor`.
|
|
190
|
-
- **TUI** — personalize this window's surroundings: sidebar header lines
|
|
191
|
-
and blocks (context/git/todo/dispatch), the Context block's detail rows
|
|
192
|
-
and gauge style (`both/bar/percent/off`), and the two Pantheon footer
|
|
193
|
-
lines. Stored in the host's durable per-plugin storage — changes apply
|
|
194
|
-
live (the sidebar re-renders on save) and survive TUI restarts.
|
|
195
|
-
|
|
196
|
-
The installer selects the `plugins` component by default; in interactive
|
|
197
|
-
setup you can change the component selection. A component-scoped install
|
|
198
|
-
includes the TUI only when `plugins` is selected. The TUI ships no quota
|
|
199
|
-
panel: for provider
|
|
200
|
-
quota and usage inside OpenCode, install the recommended community plugin
|
|
201
|
-
[`@slkiser/opencode-quota`](https://github.com/slkiser/opencode-quota) — see
|
|
202
|
-
[Quota & usage](docs/tui-v2-quota-providers.md). `pantheon_cost` and
|
|
203
|
-
`pantheon_model` keep working unchanged.
|
|
204
|
-
|
|
205
|
-
Vision requests to OpenCode Zen require a valid active session and are rejected
|
|
206
|
-
before image fetching or model-catalog lookup when the session is missing or
|
|
207
|
-
invalid. The Python vision MCP keeps its local metadata/OCR fallback when no
|
|
208
|
-
gateway credential is configured. In the plugin, an explicit Vision model
|
|
209
|
-
pin that is unavailable or ineligible returns an actionable error rather than
|
|
210
|
-
silently switching models; remove the pin to use automatic model discovery.
|
|
211
|
-
|
|
212
|
-
Memory is used selectively: the coordinator searches once only when project
|
|
213
|
-
history could affect a task, then passes relevant results to specialists. Store
|
|
214
|
-
one concise, reusable outcome at task completion when it will matter later;
|
|
215
|
-
routine child summaries stay in the handoff. See [Memory](docs/MEMORY.md).
|
|
216
|
-
|
|
217
|
-
## Compatibility
|
|
218
|
-
|
|
219
|
-
OpenCode **2.x** is supported. **OpenCode 2.0.22 is known broken.** **2.0
|
|
220
|
-
removes the OpenCode V1 contract entirely** — 1.7.x was the last line with V1
|
|
221
|
-
support.
|
|
222
|
-
|
|
223
|
-
The support policy, the known-broken exception, and the verified-version
|
|
224
|
-
evidence have exactly one home: [Compatibility](docs/COMPATIBILITY.md).
|
|
225
|
-
|
|
226
|
-
### Plugin options
|
|
227
|
-
|
|
228
|
-
The host hands a plugin its `options` only from the object form of the entry.
|
|
229
|
-
`init` writes that form for the optional per-request context pruner; memory
|
|
230
|
-
retrieval stays an explicit `memory_search` MCP call.
|
|
231
|
-
|
|
232
|
-
```json
|
|
233
|
-
{
|
|
234
|
-
"plugins": [
|
|
235
|
-
{
|
|
236
|
-
"package": "/path/to/pantheon-opencode/src/plugin",
|
|
237
|
-
"options": {
|
|
238
|
-
"context_pruner": { "enabled": true }
|
|
239
|
-
}
|
|
240
|
-
}
|
|
241
|
-
]
|
|
242
|
-
}
|
|
110
|
+
### 1. Fast-Track & Test Slicing (Micro-Edits < 10s)
|
|
111
|
+
For routine bug fixes, style adjustments, or localized refactors:
|
|
112
|
+
- **No planning ceremony:** Directly executes bounded changes.
|
|
113
|
+
- **Targeted test slicing:** Runs ONLY the test slice relevant to the modified file (e.g. `node --test tests/foo.test.mjs` or `pytest -k <func>`). Full repository test suites are reserved for CI and release gates.
|
|
114
|
+
- **Single-turn delivery:** Verifies and completes in under 10 seconds.
|
|
115
|
+
|
|
116
|
+
### 2. Assertive Council Synthesis (`/pantheon`)
|
|
117
|
+
When confronting architectural crossroads, technology evaluations, or security tradeoffs:
|
|
118
|
+
```text
|
|
119
|
+
/pantheon Should we use WebSockets or Server-Sent Events for live push notifications?
|
|
243
120
|
```
|
|
121
|
+
- **Parallel Dispatch:** Dispatches 2–3 specialist agents simultaneously with independent prompts.
|
|
122
|
+
- **Assertive Arbitration:** Rather than looping in repetitive debate rounds, divergences are resolved immediately via **Domain Authority Precedence**:
|
|
123
|
+
- **Security / Auth:** `@themis` has veto power over vulnerable designs.
|
|
124
|
+
- **Database / Schema:** `@demeter` carries authoritative weight on query performance and transaction safety.
|
|
125
|
+
- **Frontend / UX / A11y:** `@aphrodite` governs user interface responsiveness and accessibility.
|
|
126
|
+
- **Infrastructure / CI:** `@prometheus` governs deployment contracts and container isolation.
|
|
127
|
+
- **API / Protocols:** `@hermes` governs contract consistency and transport schemas.
|
|
128
|
+
- **Architecture / Modularity:** `@athena` governs separation of concerns.
|
|
129
|
+
- **JEV Multi-Criteria Decision:** If trade-offs balance competing requirements equally, Zeus invokes `pantheon_jev_decide` (or local decision matrix) to rank candidates deterministically in one tool call.
|
|
130
|
+
|
|
131
|
+
### 3. Checkpointed Deepwork (`/pantheon-deepwork`)
|
|
132
|
+
For multi-phase tasks spanning 10+ turns or multiple sessions:
|
|
133
|
+
```text
|
|
134
|
+
/pantheon-deepwork "Implement OAuth2 PKCE authentication flow"
|
|
135
|
+
```
|
|
136
|
+
- **Phase 0 (Plan):** Athena maps affected modules and returns a complete Markdown plan displayed directly in the main conversation.
|
|
137
|
+
- **Phase 1 (Discovery):** Apollo performs structured exploration.
|
|
138
|
+
- **Phase 2–N (Implementation):** Specialists implement features in parallel phases.
|
|
139
|
+
- **Phase Gates:** Themis audits each phase (`PASS` or `FAIL`).
|
|
140
|
+
- **Resumability:** Checkpoints persisted to `.pantheon/deepwork/<slug>/`. Resume anytime with `/pantheon-deepwork --resume <slug>`.
|
|
141
|
+
|
|
142
|
+
---
|
|
244
143
|
|
|
245
|
-
|
|
246
|
-
options survive a re-run. The context pruner is off unless `enabled` is
|
|
247
|
-
literally `true`.
|
|
144
|
+
## The 14 Specialized Agents
|
|
248
145
|
|
|
249
|
-
|
|
146
|
+
Every root dispatch carries a concise handoff brief — objective, bounded scope, relevant inputs, constraints, expected return format, and verification. Work is routed qualitatively — uncertainty, domain expertise, dependencies, risk, and verification — not by fixed file or line thresholds. Return contracts are role-specific: implementation uses `## subtask_summary`, investigation returns structured findings, Athena returns the complete plan, and council responses use `## specialist_response`.
|
|
147
|
+
|
|
148
|
+
| Agent | Domain Authority | Core Responsibilities |
|
|
250
149
|
|---|---|---|
|
|
251
|
-
|
|
|
252
|
-
|
|
253
|
-
|
|
254
|
-
|
|
255
|
-
|
|
256
|
-
|
|
257
|
-
|
|
258
|
-
|
|
259
|
-
|
|
260
|
-
|
|
261
|
-
|
|
262
|
-
|
|
263
|
-
|
|
264
|
-
|
|
265
|
-
documentation. Tool availability, background support, and completion
|
|
266
|
-
observability are runtime-dependent; when completion is not observable, use
|
|
267
|
-
synchronous dispatch or wait for a host notification.
|
|
268
|
-
|
|
269
|
-
Every root dispatch carries a concise handoff brief — objective, bounded scope,
|
|
270
|
-
relevant inputs, constraints, expected return format, and verification. Work is
|
|
271
|
-
routed qualitatively — uncertainty, domain expertise, dependencies, risk, and
|
|
272
|
-
verification — not by fixed file or line thresholds. Return
|
|
273
|
-
contracts are role-specific: implementation uses `## subtask_summary`,
|
|
274
|
-
investigation returns structured findings, Athena returns the complete plan,
|
|
275
|
-
and council responses use `## specialist_response`. Delegated work gets at most
|
|
276
|
-
one retry after a transient failure, and Pantheon never claims deterministic
|
|
277
|
-
fan-in or job recovery without an exposed mechanism.
|
|
278
|
-
|
|
279
|
-
The delegation model, caller/target matrix, and approval gates are documented
|
|
280
|
-
in [Architecture](docs/ARCHITECTURE.md) and
|
|
281
|
-
[Permissions, Gates & Delegation](docs/PERMISSIONS-QUALITY-DELEGATION.md).
|
|
282
|
-
The [local decision router](docs/reference/local-decision-router.md) provides a
|
|
283
|
-
default-off standalone local API (`PANTHEON_LOCAL_DECISION_ROUTER=1`), the
|
|
284
|
-
separate manual `pantheon_local_decision` advisory tool, and deterministic
|
|
285
|
-
per-turn local recommendations when the user-level `enabled` preference is on.
|
|
286
|
-
The prompt hook may add an advisory marker and narrow an existing `prompt.skills`
|
|
287
|
-
array only when identifiable skill IDs overlap the local selection; this never
|
|
288
|
-
dispatches child sessions or changes permissions or approvals. Per-turn selection
|
|
289
|
-
uses no provider egress: the formerly documented direct-provider route remains
|
|
290
|
-
historical/inactive, and legacy consent, caps, period, provider selection, or
|
|
291
|
-
credentials cannot activate it. Native child-session dispatch remains separate.
|
|
292
|
-
|
|
293
|
-
The one native egress surface is the Zeus-only `pantheon_jev_decide` tool,
|
|
294
|
-
available on hosts that expose the 2.2.5 native Decisions API; its full gate
|
|
295
|
-
chain, routes, and settlement rules are documented once in
|
|
296
|
-
[Native Decisions (Jev)](docs/reference/native-decisions.md). An opt-in second
|
|
297
|
-
consumer of that chain, the J1 risk-score shadow gate, fires one bounded
|
|
298
|
-
native decision per write/execute tool call behind `PANTHEON_J1_RISK_SCORE`
|
|
299
|
-
(off by default, zero overhead when off): the request carries no user content
|
|
300
|
-
(tool name token, scope class, input sha256 only), every failure mode skips
|
|
301
|
-
silently with a reason token, and the shadow record is the whole effect — it
|
|
302
|
-
never blocks, asks, or narrows a permission. A third opt-in consumer, the J3
|
|
303
|
-
turn-complexity escalation, runs in shadow behind `PANTHEON_J3_ESCALATION`
|
|
304
|
-
(off by default): it classifies each delegate dispatch with one bounded score
|
|
305
|
-
question (Brief sha256, length and a risk-word flag only — the Brief text
|
|
306
|
-
never leaves the machine), local rails (risk words, user-pinned models, large
|
|
307
|
-
context) and every unsure answer always favor the main model, and the recorded
|
|
308
|
-
verdict never changes the dispatched model — the per-agent override mechanism
|
|
309
|
-
is proved by test and reserved for a separately reviewed activation with its
|
|
310
|
-
own measured baseline. A fourth consumer needs no flag at all: the tool's
|
|
311
|
-
`mode: "review"` (J4 output-quality scoring) runs only when Zeus explicitly
|
|
312
|
-
asks, after a delegation completes — one typed round-trip returns an advisory
|
|
313
|
-
quality score against named criteria plus a needs-review flag. The result
|
|
314
|
-
content never leaves the machine (shape metrics and criteria only, hostile
|
|
315
|
-
criteria rejected outright), and the verdict is advisory telemetry plus a
|
|
316
|
-
flag: it never blocks, retries, or alters the delegated result. Two more
|
|
317
|
-
orchestrator-owned tool modes round out the surface: `mode: "rank"` picks the
|
|
318
|
-
best of 2–8 named candidates for a state in one typed round-trip (it returns
|
|
319
|
-
the winner and that candidate's fit score — the host API answers one scalar
|
|
320
|
-
per question, so no full ordering is returned), and `mode: "verify"` checks
|
|
321
|
-
1–4 named claims with one predicate question per claim, mapping each answer to
|
|
322
|
-
`PASS` / `FAIL` / `BLOCK` where every refusal, timeout or malformed answer is
|
|
323
|
-
`BLOCK` (fail-closed); claims are bounded caller-provided strings, hostile
|
|
324
|
-
shapes rejected outright, and neither mode adds a telemetry point.
|
|
325
|
-
|
|
326
|
-
The installer may retain a legacy target-specific preference; it does not control
|
|
327
|
-
runtime routing. `npx pantheon-opencode decision-router ...` manages stored
|
|
328
|
-
user-level preferences and budget state: `enabled` gates the local per-turn
|
|
329
|
-
recommendation only; provider-selection and budget/consent state do not activate
|
|
330
|
-
egress. The separate SQLite ledger has no default caps. Installation does not
|
|
331
|
-
contact a model provider or inspect credentials. The published CLI imports
|
|
332
|
-
committed ESM bundles:
|
|
333
|
-
`npm run build:decision-router` regenerates them and `npm run check:decision-router`
|
|
334
|
-
(the `prepack`/`prepublishOnly` drift gate) fails when any is stale. The
|
|
335
|
-
`pantheon-opencode/local-decision-router` subpath itself is TypeScript source, so
|
|
336
|
-
it loads only in the host runtime or under a transpiler such as `tsx`.
|
|
337
|
-
|
|
338
|
-
Decision-layer commands are separated by effect. `decision-router help`,
|
|
339
|
-
`status`, `preview` and `holds` are read-only: they never create, migrate, or
|
|
340
|
-
write the config or the SQLite ledger, and neither does `setup` without
|
|
341
|
-
`--apply`. `select` records a preference and can never enable anything — moving
|
|
342
|
-
the target also revokes consent recorded for the previous one. `setup` prints a
|
|
343
|
-
**computed diff of the target artifacts** (snapshot + SHA-256 per file, compared
|
|
344
|
-
against the exact bytes the writer would produce); `setup --apply` is the form
|
|
345
|
-
that writes, and a second run reports `0 create · 0 update`. It is deliberately
|
|
346
|
-
not the installer's "Change plan", which announced 112 creates for an apply that
|
|
347
|
-
created nothing. `setup --target <dir>` also reconciles the
|
|
348
|
-
`.pantheon/install-state.json` provider/consent record with the user-level config;
|
|
349
|
-
that record is derived from the config and never asserts consent the config lacks.
|
|
350
|
-
Neither path creates a backup: installer backup sets accumulate once per
|
|
351
|
-
application and contain credential-bearing material (API keys/tokens), so no
|
|
352
|
-
extra copy is made. Installer file-ownership tracking is partial and asymmetric —
|
|
353
|
-
`file_inventory` indexes 53 PROJECT artifacts and no GLOBAL ones (F-15); the
|
|
354
|
-
106-project / 107-global denominators are derived from the 213-artifact total and
|
|
355
|
-
are not separately indexed — so this preview compares the decision-layer
|
|
356
|
-
artifacts it owns, not "everything the installer touched".
|
|
357
|
-
|
|
358
|
-
The old OpenAI/TypeSafe per-turn provider adapters and endpoint descriptions are
|
|
359
|
-
retained as historical, inert code; no legacy consent, cap, period, provider, or
|
|
360
|
-
permission setting wires them into the production prompt hook. The captured-hook
|
|
361
|
-
boundary test supplies an isolated synthetic consent/provider config and fake
|
|
362
|
-
test-only credential, verifies a local recommendation marker, and verifies zero
|
|
363
|
-
native Decisions or global `fetch` calls. This is an offline regression check of
|
|
364
|
-
that hook, not a claim about unexercised network paths or a live-provider test.
|
|
365
|
-
Measured surface coverage is
|
|
366
|
-
reported honestly rather than implied: P7 runs behind `PANTHEON_P7_LOOP` with
|
|
367
|
-
an **advisory** execution cap (never described as enforcement), P2 is
|
|
368
|
-
**advisory-only** (an empirical no-op on the default path), P9 was measured on
|
|
369
|
-
top-level **and child** sessions with `reason:"overflow"` still unmeasured, and
|
|
370
|
-
P8 is **advisory: not activatable by measurement** — its declared rules match no
|
|
371
|
-
delivered permission payload, so it narrows nothing (PROBE-F).
|
|
372
|
-
|
|
373
|
-
### The decision layer is declared in full and wired in part
|
|
374
|
-
|
|
375
|
-
The decision layer declares eleven integration points (P1–P11) over three
|
|
376
|
-
decision primitives — `ask`, `rank`, and `verify` — and a deterministic policy
|
|
377
|
-
that decides which point may run. **Seven of the eleven have a production call
|
|
378
|
-
site: two behind flags that are off by default, five in shadow.**
|
|
379
|
-
|
|
380
|
-
- **P7 (loop)** is wired to `session.execution.succeeded` behind
|
|
381
|
-
`PANTHEON_P7_LOOP`, which is off. It writes each observation to an
|
|
382
|
-
in-memory, bounded telemetry collector that the plugin never reads back and
|
|
383
|
-
exposes only to tests, so "records" here does not mean an operator-visible log.
|
|
384
|
-
- **P8 (guardrail)** is wired to `permission.evaluate` behind
|
|
385
|
-
`PANTHEON_P8_GUARDRAIL`, which is off. C8 is closed **by measurement**, and
|
|
386
|
-
the measurement is negative: PROBE-F delivered 24 real `permission.evaluate`
|
|
387
|
-
events and every declared rule matched none of them (a valid `action` token and
|
|
388
|
-
a top-level `sessionID` are always present), the host resolves an unrecognised
|
|
389
|
-
effect as `ask` and proceeds, and `permission.evaluate.egress` is unreachable
|
|
390
|
-
(a tool declaring it is never advertised to the provider). As declared, P8 does
|
|
391
|
-
not narrow. The flag stays off.
|
|
392
|
-
- **P1 and P2** are registered on `session.hook('prompt')` in **shadow**. They
|
|
393
|
-
observe the same admission payload through a projection that carries only
|
|
394
|
-
`prompt.agents` and `prompt.skills`: no text, no file bytes, no metadata, no
|
|
395
|
-
mutation, no throw. Every record says `effect: none` and goes to an in-memory,
|
|
396
|
-
bounded collector the plugin never reads back (test-visible only).
|
|
397
|
-
- **P3** is wired at its declared `tool.invoke` in **shadow**: the
|
|
398
|
-
`pantheon_local_decision` tool hands the adapter its own input, whose schema
|
|
399
|
-
requires `complexity` and `gateStatus`. It records a classification when the
|
|
400
|
-
tool is invoked — never a permanent `complexity.invalid` refusal from a payload
|
|
401
|
-
that cannot carry the vector — and applies nothing.
|
|
402
|
-
- **P9** is wired at its declared `session.hook('compaction')` in **shadow**, on
|
|
403
|
-
the carrier that was actually measured: the top-level `sessionID` and the
|
|
404
|
-
`system` part array. It records the decision key and the payload shape and
|
|
405
|
-
pushes nothing — the injected compaction context is unchanged — and a payload
|
|
406
|
-
whose only session id is a nested `session.id` (a field the 2.2.5 host never
|
|
407
|
-
sends) produces no decision at all. The same shape was measured in a **child**
|
|
408
|
-
session (PROBE-G: hook fires 3/3, child `sessionID` at top level, `system` an
|
|
409
|
-
array of 4), so no child-specific branch exists. `reason:"overflow"` was not
|
|
410
|
-
inducible and is not designed for. No measured surface carries the memory
|
|
411
|
-
candidate list it would retain or discard, so its record states that absence
|
|
412
|
-
rather than a retention or a priority.
|
|
413
|
-
- **P11** is wired inside `ctx.tool.transform` in **shadow**: it reads the draft
|
|
414
|
-
the host hands that callback and calls none of its mutators. Shadow has no
|
|
415
|
-
responder, so it records the validated `choice` question over the draft ids —
|
|
416
|
-
never a selection.
|
|
417
|
-
|
|
418
|
-
**The other four — P4, P5, P6 and P10 — are library code with no production call
|
|
419
|
-
site.** Nothing constructs or registers them outside the per-point test suites —
|
|
420
|
-
the offline harness imports no adapter module. P6 and P10 declared the prompt
|
|
421
|
-
hook, but that payload never carries council candidates or counsellor claims, so
|
|
422
|
-
they stay unwired instead of recording a refusal from a surface that cannot feed
|
|
423
|
-
them. P4 ranks a memory candidate array and the measured `context` payload
|
|
424
|
-
carries `system` text parts, not candidates; P5 verifies a declared
|
|
425
|
-
`{ requirements, responses }` pair and the `execute.after` event carries
|
|
426
|
-
neither. On a host they do not read the turn, record no telemetry and have no
|
|
427
|
-
switch to turn off, because nothing registers them.
|
|
428
|
-
|
|
429
|
-
"Shadow" therefore means no behavioural change, not a claim of effect: the five
|
|
430
|
-
shadow points read a narrow slice of their surface and decide nothing from it,
|
|
431
|
-
and the two flag-gated points are off. No activation flag is set in a default
|
|
432
|
-
install.
|
|
433
|
-
|
|
434
|
-
The CLI separates effects: `npx pantheon-opencode decision-router help`, `status`,
|
|
435
|
-
`preview` and `holds` never write, `setup` writes only under `--apply`, `preview`
|
|
436
|
-
is an alias of `setup`, and `select` records a preference without enabling
|
|
437
|
-
anything. Egress stays opt-in and off, and only the explicit `enable`, `consent`,
|
|
438
|
-
`caps` and `period` actions can change that; no read-only command can.
|
|
439
|
-
|
|
440
|
-
What the layer does **not** do today, stated as measured rather than implied:
|
|
441
|
-
|
|
442
|
-
- **P2 (skills) is advisory-only and does not filter.** It is an empirical no-op
|
|
443
|
-
on the default path, because the host sends `prompt.skills` absent. Effective
|
|
444
|
-
project-skill filtering was cancelled by measurement, not emulated, and no
|
|
445
|
-
token reduction is claimed.
|
|
446
|
-
- **P7 (loop) is behind `PANTHEON_P7_LOOP`, off by default, and does not
|
|
447
|
-
enforce loop limits.** It counts executions and writes an `effect: none`
|
|
448
|
-
observation to an in-memory collector that nothing reads back. Its
|
|
449
|
-
12-execution cap is **advisory**: no measured host surface signals a new user
|
|
450
|
-
turn, so nothing is claimed to reset the streak.
|
|
451
|
-
- **P9 (compaction) is measured in top-level and child sessions.** PROBE-G fired
|
|
452
|
-
the hook in a real child 3/3 with the child's `sessionID` at top level and the
|
|
453
|
-
same payload shape. `reason: "overflow"` was **not inducible** in 6 error
|
|
454
|
-
shapes — the host retried and completed, and the one failure-recovery
|
|
455
|
-
compaction was persisted as `reason:"auto"` — so that path stays unmeasured;
|
|
456
|
-
the checkpoint reinjection inside a child was not exercised, and the `messages`
|
|
457
|
-
channel remains UNKNOWN. The memory candidate list is not available on any
|
|
458
|
-
measured surface, so the wired P9 record names that absence and claims no
|
|
459
|
-
retention or priority.
|
|
460
|
-
- **P8 (guardrail) is not activatable by measurement.** PROBE-F delivered 24 real
|
|
461
|
-
`permission.evaluate` events: the payload is
|
|
462
|
-
`["sessionID","agent","action","resources","metadata","source","effect"]` with
|
|
463
|
-
`effect` an own writable data property pre-populated by the host, and a written
|
|
464
|
-
`deny` **is** honoured. But neither declared rule can match a delivered payload
|
|
465
|
-
(valid `action` token, `sessionID` always present), the host lets an
|
|
466
|
-
unrecognised effect resolve as `ask` and execute, and
|
|
467
|
-
`permission.evaluate.egress` is unreachable — so the current rule set matches
|
|
468
|
-
nothing and does not narrow. The point is recorded as **advisory**, not as
|
|
469
|
-
awaiting review, and the inertness is the measured verdict rather than a design
|
|
470
|
-
claim. `execute.before` remains the only hard barrier.
|
|
471
|
-
- **The embeddings-versus-BM25 comparison has not been run.** The environment is
|
|
472
|
-
offline with no embedding model available, so BM25 remains the baseline and no
|
|
473
|
-
gain or loss in retrieval or tokens is claimed in either direction.
|
|
474
|
-
|
|
475
|
-
Calibration that was run: the offline harness
|
|
476
|
-
(`npx tsx tests/pantheon/decision-harness.test.ts`) executed 264 synthetic
|
|
477
|
-
cases — 132 normal, 66 refusal, 66 adversarial — across the eleven points, with
|
|
478
|
-
24 cases per point and the suite asserting at least 20 cases, 5 refusals and 5
|
|
479
|
-
adversarial cases per point. All 264 cases passed on 2026-10-31. This harness is
|
|
480
|
-
the only place all eleven points are exercised in one run, but it imports no
|
|
481
|
-
adapter module: it iterates the registry and exercises the shared primitives,
|
|
482
|
-
while the adapters are covered by their per-point suites
|
|
483
|
-
(`decision-p1.test.ts` … `decision-p11.test.ts`). Passing it says nothing
|
|
484
|
-
about whether a point is registered in a host.
|
|
485
|
-
The harness measures contracts (admission, refusal, shadow no-op, telemetry
|
|
486
|
-
redaction, offline discipline, registry integrity), **not** token use, quality,
|
|
487
|
-
latency or any effect inside a host turn. Numbers and unmeasured surfaces are
|
|
488
|
-
recorded in the [router reference](docs/reference/local-decision-router.md).
|
|
150
|
+
| **@zeus** | Central Orchestrator | Directs workflow, manages delegations, arbitrates council decisions, enforces token efficiency. Never implements directly. |
|
|
151
|
+
| **@athena** | Architecture & Planning | Research-first planning, system boundaries, LTS policy, dependency analysis. Plan-only; never implements. |
|
|
152
|
+
| **@hermes** | Backend & APIs | FastAPI, Python async, Node/TypeScript backends, API contracts, protocol conformance. |
|
|
153
|
+
| **@aphrodite** | Frontend & UX | React 19, strict TypeScript, responsive design, WCAG accessibility, CSS optimization. |
|
|
154
|
+
| **@demeter** | Database & Persistence | Schema migrations, query optimization, N+1 query prevention, transaction integrity. |
|
|
155
|
+
| **@themis** | Security & Quality Gate | OWASP audits, static analysis (Biome/ruff), authorization gates, dead code removal. Reviews material changes. |
|
|
156
|
+
| **@prometheus** | DevOps & Infrastructure | Docker, CI/CD pipelines, runtime configuration, provider abstraction, telemetry infra. |
|
|
157
|
+
| **@talos** | Hotfix Fast Lane | Rapid, standalone fixes for small bugs, typos, and minor logic without subagent overhead. |
|
|
158
|
+
| **@apollo** | Investigation Scout | Read-only discovery: parallel codebase and web searches. Returns structured findings. |
|
|
159
|
+
| **@hephaestus**| AI Pipelines & RAG | Embeddings, vector stores, prompt engineering, agent pipelines. |
|
|
160
|
+
| **@gaia** | Geospatial & Sensing | Remote sensing, SAR, geospatial analysis, ML/DL raster classification (read-only). |
|
|
161
|
+
| **@iris** | Git & Release Ops | Branching, PRs, issue triage, release notes. Never pushes or merges without human approval. |
|
|
162
|
+
| **@mnemosyne** | Project Memory Bank | Maintains `.pantheon/memory-bank/` and records ADRs when explicitly requested. |
|
|
163
|
+
| **@nyx** | Observability & Telemetry| OpenTelemetry tracing, token/cost monitoring, execution profiling. |
|
|
489
164
|
|
|
490
|
-
|
|
165
|
+
---
|
|
491
166
|
|
|
492
|
-
|
|
493
|
-
|---|---|
|
|
494
|
-
| [Installation](docs/INSTALLATION.md) | [Architecture](docs/ARCHITECTURE.md) |
|
|
495
|
-
| [Quick Start](docs/QUICKSTART.md) | [MCP](docs/MCP.md) · [MCP tools](docs/mcp-tools.md) |
|
|
496
|
-
| [Upgrading](docs/UPGRADING.md) | [Memory](docs/MEMORY.md) |
|
|
497
|
-
| [Compatibility](docs/COMPATIBILITY.md) | [OpenCode 2.x](docs/opencode.md) |
|
|
498
|
-
| [Agents & Skills](docs/agents/README.md) | [Platforms](docs/PLATFORMS.md) |
|
|
499
|
-
| [Release Process](docs/RELEASING.md) | [Branch Protection](docs/BRANCH-PROTECTION.md) |
|
|
500
|
-
| [Model Routing](docs/reference/model-routing.md) | [Local Decision Router](docs/reference/local-decision-router.md) · [Native Decisions (Jev)](docs/reference/native-decisions.md) · [Offline Evaluation Protocol](docs/reference/offline-evaluation.md) · [Env Vars](docs/reference/env-vars.md) · [Code Mode](docs/reference/code-mode.md) |
|
|
501
|
-
| [Contributing](CONTRIBUTING.md) | [Roadmap](ROADMAP.md) · [Changelog](CHANGELOG.md) |
|
|
502
|
-
|
|
503
|
-
Memory writes through `pantheon-memory` are serialized, and failed writes are
|
|
504
|
-
reported as failures rather than as stored entries. Eval resources wait for
|
|
505
|
-
SQLite locks and close their connections; project-specific resources resolve
|
|
506
|
-
the current project root per request while preserving the trusted
|
|
507
|
-
`PANTHEON_HOME` fallback. See [Memory](docs/MEMORY.md) for store separation and
|
|
508
|
-
[MCP tools](docs/mcp-tools.md) for details.
|
|
509
|
-
|
|
510
|
-
Native-MCP usage is measured, not assumed: `node bench/mcp-usage-audit.mjs`
|
|
511
|
-
writes a dated audit of real memory/persistence/vision/code-mode/resource
|
|
512
|
-
usage (reports stay local — they carry bounded session excerpts), and the
|
|
513
|
-
agent usage policies derived from it are documented in
|
|
514
|
-
[MCP tools](docs/mcp-tools.md).
|
|
515
|
-
|
|
516
|
-
The release install gate's tool canary loads the installed tarball's plugin
|
|
517
|
-
entry under OpenCode 2.0.30 and verifies `hashline_edit` registration, Code Mode
|
|
518
|
-
visibility, and a bounded edit in an isolated project. It does not prove npm
|
|
519
|
-
publication or compatibility with untested host versions; details are in the
|
|
520
|
-
[release process](docs/RELEASING.md).
|
|
521
|
-
|
|
522
|
-
Generate beta release notes from commits since the latest prior beta tag with
|
|
523
|
-
`node scripts/versioning.mjs apply --beta --notes`. See the [release
|
|
524
|
-
procedure](docs/RELEASING.md) for details.
|
|
525
|
-
After the version bump, verify and rebuild the TUI workspace from the repository
|
|
526
|
-
root with `npm run typecheck --workspace=src/plugins/tui` and
|
|
527
|
-
`npm run build --workspace=src/plugins/tui`.
|
|
528
|
-
The maintainer-only `release` scope is excluded from public notes and their
|
|
529
|
-
Closed Issues references; see the [release-notes policy](docs/RELEASING.md#release-notes).
|
|
530
|
-
|
|
531
|
-
## Status
|
|
532
|
-
|
|
533
|
-
The manifests in this checkout express the version being prepared, not a
|
|
534
|
-
confirmed publication. Check the npm `beta` dist-tag for what is actually
|
|
535
|
-
published. See [Compatibility](docs/COMPATIBILITY.md) for the supported hosts
|
|
536
|
-
and [releases](https://github.com/ils15/pantheon-opencode/releases) for the
|
|
537
|
-
latest published changes.
|
|
538
|
-
|
|
539
|
-
## What's new
|
|
540
|
-
|
|
541
|
-
See the [GitHub release](https://github.com/ils15/pantheon-opencode/releases)
|
|
542
|
-
and [CHANGELOG.md](CHANGELOG.md) for the changes in 1.7.0.
|
|
543
|
-
|
|
544
|
-
## Contribute
|
|
545
|
-
|
|
546
|
-
Ideas, bug reports, documentation improvements, and code contributions are
|
|
547
|
-
welcome. Read [CONTRIBUTING.md](CONTRIBUTING.md) before opening an issue or pull
|
|
548
|
-
request. That file also defines the documentation rule: one home per fact, with
|
|
549
|
-
a short list of facts that must not be duplicated.
|
|
550
|
-
|
|
551
|
-
For code changes, run the smallest relevant check and expand it to match the
|
|
552
|
-
change's risk. `npm test` is the full repository suite; see
|
|
553
|
-
[CONTRIBUTING.md](CONTRIBUTING.md) for when to use it.
|
|
554
|
-
|
|
555
|
-
## Citation and DOI
|
|
556
|
-
|
|
557
|
-
Pantheon is released under the [MIT License](LICENSE). Cite the
|
|
558
|
-
[Zenodo concept DOI](https://doi.org/10.5281/zenodo.22650136), which always
|
|
559
|
-
resolves to the latest archived release; each release also has its own version
|
|
560
|
-
DOI. Citation metadata is in [CITATION.cff](CITATION.cff).
|
|
167
|
+
## Skills & MCP Tooling
|
|
561
168
|
|
|
562
|
-
|
|
169
|
+
- **In-Process Skill Search (`pantheon_skill_search`):** Quickly discovers local skills and external ecosystem skills without requiring shell execution permissions.
|
|
170
|
+
- **25 Production Skills:** Includes `test-slicing`, `fast-path-execution`, `tdd-with-agents`, `instruction-forge`, and `incremental-implementation`.
|
|
171
|
+
- **5 Local MCP Servers:**
|
|
172
|
+
- `pantheon-memory`: Shared project memory with atomic SQLite locking.
|
|
173
|
+
- `pantheon-persistence`: File and context persistence.
|
|
174
|
+
- `pantheon-resources`: Local documentation and asset serving.
|
|
175
|
+
- `pantheon-code-mode`: Sandboxed Python execution.
|
|
176
|
+
- `pantheon-vision`: Image analysis and diagnostic visual inspection.
|
|
177
|
+
- **Advisory JEV Decision Router:** Evaluates candidate options (`pantheon_jev_decide`), orders trade-offs, and provides formal multi-criteria decision making without modifying permissions.
|
|
563
178
|
|
|
564
179
|
---
|
|
565
180
|
|
|
566
|
-
|
|
181
|
+
## Settings & TUI Window (`/pantheon-settings`)
|
|
182
|
+
|
|
183
|
+
Run `/pantheon-settings` (or `Pantheon: Settings` in the command palette) to open the interactive settings window:
|
|
184
|
+
|
|
185
|
+
- **Gates & Overrides:** Toggle named behaviors live; persists to `pantheon-overrides.json`.
|
|
186
|
+
- **Model Presets:** Switch active model tier globally or per project.
|
|
187
|
+
- **Decision Model:** Configure local and remote JEV decision routing.
|
|
188
|
+
- **Updates:** Check registry channel (`beta` / `stable`) and trigger upgrades.
|
|
189
|
+
- **Doctor:** Run inline system health diagnostics.
|
|
190
|
+
- **TUI Personalization:** Adjust spacing (`comfortable` vs `compact`), separators, gauge bar styles, and context display rows.
|
|
191
|
+
- **Non-Destructive Overlay & Mouse Navigation:** Opens as a floating dialog overlay over the active session without blanking the workspace (configurable via `dialog` vs `fullscreen`), with full mouse click support for tabs, sidebar, rows, value cyclers, and actions.
|
|
192
|
+
- **Modernized What's New Window:** Card layout with rounded border, pagination buttons, and a dedicated `[OK, Entendido]` button with keyboard (`Enter`/`Esc`/`Q`) and mouse click support.
|
|
193
|
+
|
|
194
|
+
---
|
|
195
|
+
|
|
196
|
+
## Documentation
|
|
197
|
+
|
|
198
|
+
| Start Here | Architecture & Deep Dives | Reference & Operations |
|
|
199
|
+
|---|---|---|
|
|
200
|
+
| [Quick Start](docs/QUICKSTART.md) | [Architecture](docs/ARCHITECTURE.md) | [MCP Tools](docs/mcp-tools.md) |
|
|
201
|
+
| [Installation Guide](docs/INSTALLATION.md) | [Delegation & Gates](docs/PERMISSIONS-QUALITY-DELEGATION.md) | [Native Decisions (Jev)](docs/reference/native-decisions.md) |
|
|
202
|
+
| [Compatibility](docs/COMPATIBILITY.md) | [Memory System](docs/MEMORY.md) | [Model Routing](docs/reference/model-routing.md) |
|
|
203
|
+
| [Upgrading](docs/UPGRADING.md) | [OpenCode 2.x Integration](docs/opencode.md) | [Changelog](CHANGELOG.md) |
|
|
204
|
+
|
|
205
|
+
---
|
|
206
|
+
|
|
207
|
+
## What's New in v2.0.0-rc.1
|
|
208
|
+
|
|
209
|
+
- **Release Candidate Series Inauguration:** Production-ready multi-agent orchestration for OpenCode 2.x.
|
|
210
|
+
- **Agent Token Compaction (-44.3%):** Removed duplicated backend code in Apollo/Prometheus, saving 8,150 tokens per turn (~29.3M tokens/month).
|
|
211
|
+
- **Pragmatism & Anti-Pedantry Directive:** Safe diagnostic operations are authorized without bureaucratic gates, ending agent sluggishness.
|
|
212
|
+
- **Batched Suite TDD & Test-Slicing:** Cohesive multi-endpoint testing and sub-100ms targeted test execution.
|
|
213
|
+
- **Proactive Skill Discovery & Query Caching:** Cached skill lookups (24h TTL) and automated recommendation for external tools.
|
|
214
|
+
- **Live JEV Scorecard & Sandbox Web App:** 16/16 pass rate ($0.0000238/decision) verified with an end-to-end sandbox application.
|
|
215
|
+
|
|
216
|
+
See [CHANGELOG.md](CHANGELOG.md) for full historical release notes.
|
|
217
|
+
|
|
218
|
+
## Contributing
|
|
219
|
+
|
|
220
|
+
Ideas, bug reports, and contributions are welcome! Please read [CONTRIBUTING.md](CONTRIBUTING.md) before submitting a PR.
|
|
221
|
+
|
|
222
|
+
## Citation and License
|
|
223
|
+
|
|
224
|
+
Pantheon is open-source software licensed under the [MIT License](LICENSE).
|
|
225
|
+
Cite via Zenodo concept DOI: [10.5281/zenodo.22650136](https://doi.org/10.5281/zenodo.22650136). Citation metadata is available in [CITATION.cff](CITATION.cff).
|
|
226
|
+
|
|
227
|
+
Canonical repository: <https://github.com/ils15/pantheon-opencode>
|