|
1
|
+ |
# Coder push handoff
|
|
2
|
+ |
|
|
3
|
+ |
Date: 2026-08-24
|
|
4
|
+ |
|
|
5
|
+ |
Status: handoff. Written when the session that landed the work below paused.
|
|
6
|
+ |
The next agent should read this first, then the three planning documents it
|
|
7
|
+ |
points at.
|
|
8
|
+ |
|
|
9
|
+ |
## 1. Read these first
|
|
10
|
+ |
|
|
11
|
+ |
| Document | What it decides |
|
|
12
|
+ |
| --- | --- |
|
|
13
|
+ |
| `docs/2026-08-24-registry-network-strategy.md` | Why the plugin registry is the network, why the cloud is centralized, why consent is load-bearing. |
|
|
14
|
+ |
| `docs/2026-08-24-coder-first-cloud-complements.md` | The product plan: the coder wedge, the three-lane compute mix, the T3-derived thread-plane protocol, the delivery sequence with issue numbers. |
|
|
15
|
+ |
| `docs/2026-08-24-triage-and-plugin-model-assessment.md` | The plugin contract (packet ABI, manifest, capability tiers) and the 2026-08-24 backlog triage. |
|
|
16
|
+ |
| `docs/2026-08-24-api-v1-rename-audit.md` | The `/api/v3` → `/api/v1` inventory and sequencing. |
|
|
17
|
+ |
| `docs/plugins/...` in the monorepo | `2026-08-24-coder-plugin-demo-shape.md`, the plugin host's own shape notes. |
|
|
18
|
+ |
|
|
19
|
+ |
Boards: [Coder v1](https://openagents.com/OpenAgentsInc/openagents.com/projects/12)
|
|
20
|
+ |
(project 12) and [Plugin registry](https://openagents.com/OpenAgentsInc/openagents.com/projects/13)
|
|
21
|
+ |
(project 13). Both carry cross-repository items; `openagents` numbers in this
|
|
22
|
+ |
document are the monorepo, unprefixed numbers are `openagents.com`.
|
|
23
|
+ |
|
|
24
|
+ |
## 2. What landed
|
|
25
|
+ |
|
|
26
|
+ |
Server (`openagents.com`, through `7b8f939`):
|
|
27
|
+ |
|
|
28
|
+ |
- Thread plane: web viewer at `/threads` and `/threads/{id}` with the
|
|
29
|
+ |
attach-before-snapshot live protocol (#201); events API returning the
|
|
30
|
+ |
created event with machine codes and atomic batch append (#208); grant
|
|
31
|
+ |
re-mint at `POST /api/v3/threads/{id}/grants`; an optional bounded
|
|
32
|
+ |
`repository` on threads with API filtering (#210); admission and mint
|
|
33
|
+ |
serialized on the owner row (#195).
|
|
34
|
+ |
- Providers: reasoning as a first-class `ProviderEvent` member and faithful
|
|
35
|
+ |
tool-call replay through the inference proxy (#164 server half); a typed
|
|
36
|
+ |
model catalog at `GET /api/v3/models` with loud refusal of unserved or
|
|
37
|
+ |
mismatched models and effective-model attribution (#199, closing #160's
|
|
38
|
+ |
defect class) under the new `PROVIDER-002` invariant.
|
|
39
|
+ |
- Integrity: thread grant usage on the leaderboard (#204); issuer-key
|
|
40
|
+ |
retirement refused when it would unverify signed history (#191); the
|
|
41
|
+ |
machine pairing vault given its own key under `VAULT-001` (#192); the
|
|
42
|
+ |
runbook `rebuild/1` correction plus a test that fails when any doc names a
|
|
43
|
+ |
function that does not exist (#189).
|
|
44
|
+ |
|
|
45
|
+ |
CLI (`openagents` monorepo, through `38625181fc` plus later commits from a
|
|
46
|
+ |
concurrent session):
|
|
47
|
+ |
|
|
48
|
+ |
- The coder writes its transcript to `thread_events` as the turn loop runs
|
|
49
|
+ |
(#23) and resumes a thread with `--resume` (#24), replaying both the UI
|
|
50
|
+ |
transcript and the model wire history.
|
|
51
|
+ |
- Reasoning is parsed and rendered, and the client tool loop replays faithful
|
|
52
|
+ |
`tool_calls` history (#31, completing #164 in code).
|
|
53
|
+ |
- Plugins: a working `/plugin load` path, then the walking skeleton — an
|
|
54
|
+ |
owned Rust PDK, an engine abstraction, enforced read-only mounts, and
|
|
55
|
+ |
`packet-v0` ABI pinning (#26, partially done). ATIF exports carry plugin
|
|
56
|
+ |
lifecycle and per-call provenance (#32).
|
|
57
|
+ |
- `openagents trace` discovers, summarizes, and redacts local traces; upload
|
|
58
|
+ |
refuses honestly until the server route exists (#14).
|
|
59
|
+ |
|
|
60
|
+ |
## 3. Where to pick up
|
|
61
|
+ |
|
|
62
|
+ |
In order. The first three are the Coder v1 critical path.
|
|
63
|
+ |
|
|
64
|
+ |
1. **Finish #26** (monorepo): both-sides schema validation, the `tool.ran`
|
|
65
|
+ |
receipt event on plugin runs, `/plugin list` and `/plugin unload`, and
|
|
66
|
+ |
memory-ceiling enforcement (declared-only until a wasmtime engine sits
|
|
67
|
+ |
behind the seam).
|
|
68
|
+ |
2. **#206** — the registry on the forge: plugins as repositories with typed
|
|
69
|
+ |
manifests and digest-pinned releases, a typed index, and the capability
|
|
70
|
+ |
gap loop that files issues when a search finds nothing.
|
|
71
|
+ |
3. **#217** — `POST /api/v3/traces`, the ingest route `openagents trace
|
|
72
|
+ |
upload` already names in its refusal.
|
|
73
|
+ |
4. **#212–#216** — the `/api/v1` rename. #212 (rename plus a transparent
|
|
74
|
+ |
`/api/v3` rewrite plug) wants a solo window: it touches roughly 1,530
|
|
75
|
+ |
lines across the server and should not race other branches. #215 (the gh
|
|
76
|
+ |
posture decision) gates #216.
|
|
77
|
+ |
5. **#202, #205, #77** — durable effect outbox, consent tiers, chat-to-issue
|
|
78
|
+ |
capture. Each had an agent in flight at pause; check for an unmerged
|
|
79
|
+ |
branch before restarting one (section 5).
|
|
80
|
+ |
6. **#28** (monorepo) — fleet rendering over the task registry. A concurrent
|
|
81
|
+ |
session was actively building `coder-tasks.ts`; coordinate before touching
|
|
82
|
+ |
it.
|
|
83
|
+ |
7. **#197** — the load-sensitive flaky tests. Run this alone on a quiet
|
|
84
|
+ |
machine; its whole method is hammering tests that fail under parallel
|
|
85
|
+ |
load, so it poisons other agents' runs.
|
|
86
|
+ |
|
|
87
|
+ |
Owner decisions blocking work, not agent work: **#209** (thread lifecycle —
|
|
88
|
+ |
the CLI revokes on clean exit, so resume serves only crashes and other
|
|
89
|
+ |
machines today), **#193** (no encrypted Ecto columns), **#29** (the wallet
|
|
90
|
+ |
rail), **#207** (the settlement proof's treasury spend), **#215** (gh
|
|
91
|
+ |
posture).
|
|
92
|
+ |
|
|
93
|
+ |
## 4. How to work here
|
|
94
|
+ |
|
|
95
|
+ |
- **Push to the forge, never GitHub**: `git push openagents HEAD:main`. Both
|
|
96
|
+ |
repositories.
|
|
97
|
+ |
- **Use worktrees.** The monorepo checkout is frequently dirty with another
|
|
98
|
+ |
session's work. Branch from `origin/main` into a worktree, rebase there,
|
|
99
|
+ |
and push from the worktree if the main checkout is busy. Never stash or
|
|
100
|
+ |
reset someone else's work.
|
|
101
|
+ |
- **The monorepo's pre-push gate regenerates artifacts.** After adding files,
|
|
102
|
+ |
run `pnpm run generate:assure-repo` and `pnpm run audit:assure-repo` and
|
|
103
|
+ |
commit the results, or the push is refused.
|
|
104
|
+ |
- **Server tests**: prefer targeted files. `mix precommit` runs the whole
|
|
105
|
+ |
suite (about 4,170 tests, three minutes) and is worth it before pushing
|
|
106
|
+ |
anything that touches the provider or chat lanes.
|
|
107
|
+ |
- **Invariant Discipline**: a change that adds, relaxes, or reinterprets an
|
|
108
|
+ |
invariant updates `INVARIANTS.md` in the same change, with a test.
|
|
109
|
+ |
- **Owner actions go in `NEEDS_OWNER.md`** at the workspace root, committed
|
|
110
|
+ |
and pushed — the owner reads it on GitHub, so an uncommitted edit is
|
|
111
|
+ |
invisible.
|
|
112
|
+ |
|
|
113
|
+ |
## 5. Loose ends at pause
|
|
114
|
+ |
|
|
115
|
+ |
- Five agents were in flight: #205, #202, #77 (`openagents.com`), #12
|
|
116
|
+ |
(monorepo, operator deployment commands), and the #198 foreign-session
|
|
117
|
+ |
plugin pilot (monorepo). Their branches, if they exist, are
|
|
118
|
+ |
`thread-visibility-tiers`, `effect-outbox`, `capture-issue-from-chat`,
|
|
119
|
+ |
`operator-deploy-commands`, and `foreign-sessions-plugin`, in worktrees
|
|
120
|
+ |
under the session scratchpad. Check `git branch --list` in both
|
|
121
|
+ |
repositories before redoing any of that work.
|
|
122
|
+ |
- **#164 and #160 are code-complete on main but stay open** until production
|
|
123
|
+ |
serves the revision; #187 tracks the gap between `main` and the deployed
|
|
124
|
+ |
forge. Verify against the live API before closing either.
|
|
125
|
+ |
- The published CLI is 0.3.5 and predates every coder change above. Nothing
|
|
126
|
+ |
in the coder lane reaches a user until it publishes.
|