skills: OpenCode's and Goose's folders are read too (on top of #1410) by AbirAbbas · Pull Request #1409 · Agent-Field/CodeAF

AbirAbbas · 2026-09-23T16:15:44Z

Skills (#1277), on top of #1410. It is based on santos/dev2, which already has Santosh's skills work: skills from other tools read in place, /skill, use_skill, and skills chosen per message. #1396 is still landing there. When #1410 merges and santos/dev2 is deleted, GitHub retargets this pull request to dev.

Draft. Do not merge before #1410.

What is in it now

.opencode/skills and .goose/skills are read too, under the project and the home directory. Use skills from other agent harnesses #1277's day-one list names both. They rank after the other six folders, so an earlier folder's copy of a name wins and the later copy stays in the result, marked shadowed.
A change entry.

Checked

make pr-ready BASE=origin/santos/dev2: exit 0.
Real binary, real model (deepseek-v4-flash), throwaway home, driven in tmux. Four skills were placed:
- with the npx skills CLI, patched with a codeaf entry (List codeaf in npx skills (vercel-labs/skills) once #1277 ships #1399): one in the project (add -a codeaf, lands in .agents/skills) and one global (add -g -a codeaf, lands in ~/.agents/skills)
- by hand: one in the project's .opencode/skills and one in the home's .goose/skills
All four were listed in the /skill picker. Each was used by the model on its own, and each answer carried that skill's unique marker line. The global one went through use_skill get.
On santos/dev2 as it stands, attaching from /skill does not work in the real binary. Every row says memory is off, so this cannot be attached even with memory.enabled on. The picker asserts SkillFacts on the surface's memory seam, but the launch hands it the v3Brain adapter (cmd/codeaf/chatv3.go), which does not forward that method. The picker test passes only because its fake memory has the method. Skills from Claude Code plugins and Codex, with memory off, chosen by what they are for #1396 moves that read onto the session. I built Skills from Claude Code plugins and Codex, with memory off, chosen by what they are for #1396's head with this pull request's commit cherry-picked (it applies cleanly and internal/skills tests pass). There the warning is gone, enter attaches, and the turn shows · skills · goose-honk, open-sesame, house-greeting.

Still to do here

After Skills from Claude Code plugins and Codex, with memory off, chosen by what they are for #1396 lands in santos/dev2: rebase, and add OpenCode and Goose to its skills-from-other-tools manual page in the same commit.
Use skills from other agent harnesses #1277's plan items that santos/dev2 lands on dev: the worker harness as default, runs as work, skills from other harnesses, the chat as manager #1410 does not have: per-skill slash commands, /skill add and /skill remove from a folder or a git URL, and a Skills tab. The plan's per-scope on/off conflicts with santos/dev2 lands on dev: the worker harness as default, runs as work, skills from other harnesses, the chat as manager #1410's choice by meaning, and needs Santosh's call first. The sub-issues Skills 2/6: /<skill-name> invokes a skill by its own name #1404–Skills 6/6: skills inside tasks #1408 are being re-scoped as gaps against santos/dev2 lands on dev: the worker harness as default, runs as work, skills from other harnesses, the chat as manager #1410.

🤖 Generated with Claude Code

…ends in the dots

…s the dot row Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…ail's held row keeps its title PlanTasks compared a task's id against the word `root`, but a run the conversation opens is rooted at the task's own number, so no real run's row carried progress and the dot row never drew. It asks the store for its root. On a rail under thirty cells a held row's tail took the line and left the title one letter; the title is laid first there, and yields only where it still reads beside the whole tail. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…s own edge The rail's projection kept the indent of a conversation row it never draws, so the run's root stood one level in and the page's one-level budget at that width drew a grandchild at its parent's indent. The projection draws no conversation row, affords two levels, and keeps the rail's one-cell edge. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…dots, and the narrow rail draws the dot row under the run's title The live line under a task and the task under a task each left a blank in the family's column, so the tree read as loose pieces; the walk now carries a rule through every ancestor that still has a sibling to come. A run with a failure was drawn in words alone at the widest tier, the one run most worth a picture. On a rail under 40 columns the dot row stands on the line under the title. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…unning cell keeps the frontier Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…ailed is stopped Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…own done is never cancelled Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…for it (#1207) Under CODEAF_TASK_BELT=bash a review check resolved to the mastermind tier, so the crew's checker was never asked and the review round was billed at the thinker's price. SeatFor now seats a check on the careful work tier. The door resolves the check seat in order: --check-model, CODEAF_CHECK_MODEL, the plan seat when the person pinned it by flag or by environment, else the profile's careful row. A pinned two-model run stays two-model. The check seat stays read-only.

…en questions The stamp and the rows covered every task the conversation's filter admitted, and the held questions were read from a store kind nothing writes, so they were always empty. The family is the root and what reaches it through its parents; the questions are the session's open questions whose subject works one of those tasks, by their heads. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

#1373) * tui3: the stopped frame is compared against a clock in the test's hand The head draws the time of day on every frame, so a test comparing two whole frames was comparing the wall clock too, and failed whenever its two captures straddled a minute boundary. One failure in eighty one runs, read once on a pull request as a regression in a change that could not reach the path. stoppingApp now holds its own clock, the way stopbound_test.go's boundedStopApp already does. The seam existed: app.now reads a.clock and falls back to time.Now. It is pinned there and not in newTestApp, whose calling files are in the hundreds and many of which assert on elapsed durations. Two repairs produce the same green here and only one of them is right, so the comparison has a name now and a guard on it: narrowing stoppedFrame to drop the head leaves the original test passing and turns the guard red. * docs: change entry for the stopped frame and its held clock --------- Co-authored-by: codeaf <agentfield-bot@users.noreply.github.com>

…#1374) The guard is a guard only because it calls the same function the real test calls. The comment said the helper must stay the whole frame and that narrowing it turns the guard red. It did not say why the helper exists at all, so inlining it back into its callers reads like removing a pointless indirection, leaves every test green, and detaches the guard from the comparison it guards in the same stroke. Comment and change entry only. No code changes. Co-authored-by: santoshkumarradha <eng@agentfield.ai>

… unbuilt (#1382) A person with eighty-one SKILL.md folders on disk asked whether the chat could use skills and was told codeaf has no such mechanism. The binary carried the wave. memory.enabled was off. Every part behaved as designed. The launch import returns when memory is nil, the catalog renders zero bytes for an empty shelf, and use_skill is withheld rather than present and refusing. The model was left with no shelf, no verb and no sentence about either, so it reasoned from the silence and denied a shipped feature: correctly from the inside, wrongly about the world. Absent-not-broken stops a model planning a reply around a call that can only refuse. It does not stop a model reasoning from the absence. So it gets a companion clause rather than an exception: when a capability is withheld by a setting, something has to say so. The catalog renders one line when the machine holds skills this session cannot reach, naming the setting and saying they exist. The door measures that once at launch, because a prompt prefix may not walk six folders per render. A machine with no skills still pays nothing, which is what the emptiness law is about. The /skill picker reads the folders rather than the shelf, so it listed skills that could not be attached. Each row now carries the reason beside the folder's own warning rather than instead of it. Both tests were run red against the unfixed tree before they were kept. Co-authored-by: santoshkumarradha <eng@agentfield.ai>

dev carries #1332, #1108, #1333, #1107, #1208, #1071, #1376, #1377, #1381, #1385, #1386, #1387 and #1384. #1108 is the santos/dev slice held at d39a4e3, which dev2 already descends from, so the tree was merged against d39a4e3 as its base: the slice counts as already here, and only what dev did beyond it is brought in (the nested-module walk, the image guard's duplicate, and everything from #1333 on). The commit still records origin/dev as its second parent, so a dev2 to dev pull request sees dev as merged. Four files conflicted: - tools_image.go: both sides restored the empty-prompt guard; dev2's comment kept, and dev's second copy of the guard's test dropped because #1372's test pins the same refusal. - surfacedoors_law_test.go: dev2 added one ledger door and #1071 crossed workingNowAgent; the ratchet counts the merged ledger, 22. - tui_e2e_test.go, tuiwords_test.go: the home that #1071 draws is the one in the tree (dev2 never redrew home after the slice), so the suite waits for #1071's words.

…ests do not read the shell's telemetry switch A parked parent was woken by launchWaits as soon as its children's rows read done, before their returns were absorbed. The woken parent then wrote its ending ahead of the review the child's return seats beneath it, the store refused that ending over the open check, and the run answered incomplete. review_order_test.go's parked root failed this way under a loaded box on dev2 before this merge. The wake now waits for every settled task it names to have landed, which is the rule landed() already states. Since dev's #1387 the telemetry off switch quiets the Model Pool to `read`, so a shell exporting CODEAF_TELEMETRY=off turned every pool test that means the default into a failure. cmd/codeaf's TestMain clears the switch (a test binary never reports, and the pool's submit address is already pinned), and the config row test pins it for itself.

TestAStandingFiringReachesTheHostedConversation (from dev's #1377) opened the lane through WatchTaskUpdates, whose ask is made off the loop, and then repeated the ask synchronously as its receipt. When the door's own ask reached the engine after the firing, it replaced the far subscription the firing had gone down and the row was lost. Under parallel load this failed on dev and on this branch alike. The test now installs the client lane and makes the one ask itself, answered before the firing.

# Conflicts: # internal/manual/chat/services.md

The wrapper names the lock after the suite starts, so the suite's pid on the pipe says nothing about the file. The fixture's 50 ms sleep stood in for that wait, and on a loaded box the test read an empty lock file. It now reads the file until the line is there, under a ten second deadline.

…1395) Twelve review findings on the task page, the tab mark, the pool judge sweep and the manual. Full make check on Spark at 635c029: EXIT 0.

… and through every door (#1393) Nine review findings on the chat manager, the digest, config keys and the commit trailer. Full make check on Spark at 449e792: EXIT 0.

…it (#1411) Ten review findings on the run lifecycle: no new request adopts an earlier run, every ending is written on the root, the dollar limit ends peers, close lands nothing, steers on ended runs are refused, interrupted rows are honest. Full make check on Spark at 888911c: EXIT 0.

santoshkumarradha · 2026-09-23T17:00:35Z

Heads-up on overlap with santos/dev2 → dev (#1410, going ready shortly).

internal/skills: this PR carries Skills from Claude Code plugins and Codex, with memory off, chosen by what they are for #1396's discovery (plugins.go, Codex .system). dev2 has the same package. Whichever lands second should take the other's copy whole rather than merge it line by line.
/skill and the shelf: dev2 has the first Use skills from other agent harnesses #1277 design: a memory-backed shelf, use_skill, per-message automatic choice, and a /skill picker that attaches (WIP review needed: skills tranche 1 (store, attachment, shelf, use_skill) #1327, Hosted chat says it stopped while the run keeps running #1345–Spend surfaces disagree after an interrupted streaming turn #1348, Dropping a folder shows Add but does not attach it #1380, chat, session: skills that are switched off say so instead of looking unbuilt #1382). Your steps 2–5 replace it, and the manual pages use-skill.md, putting-a-skill-in-front.md and skills-a-turn-used.md collide.
Proposal: santos/dev2 lands on dev: the worker harness as default, runs as work, skills from other harnesses, the chat as manager #1410 lands first, then this branch merges dev and removes dev2's shelf-based skills in favour of the settled plan. We've stopped extending the shelf design on dev2.

Two findings from our testing to carry into your steps:

With memory.enabled off, dev2's skills are dead, so it's good that the settled design doesn't need the store.
Automatic pickup by shared words missed reworded requests. Giving the model each turned-on skill's name and description (your step 3) is the right shape. It's worth a small check with reworded prompts.

+1 on ~/.agents/skills for #1399. ~/.codeaf/skills holds the resident's bin/ shelf.

santoshkumarradha · 2026-09-23T17:29:26Z

Proposal so skills land on dev once, in your design:

We remove our side. Right after santos/dev2 lands on dev: the worker harness as default, runs as work, skills from other harnesses, the chat as manager #1410 lands, we open a PR to dev that deletes dev2's shelf-based skills: use_skill, the store shelf, per-message automatic choice, /skill attach, and their three manual pages. internal/skills discovery stays as the shared base. Merge dev after that and you're on a clean tree.
Skills from Claude Code plugins and Codex, with memory off, chosen by what they are for #1396 closes unmerged as the reference. Its evidence, for your steps 3 and 6:
- Real installs: skills from npx skills add, claude plugin install example-skills@anthropic-agent-skills and a hand-made ~/.codex/skills skill, with memory off and on. Exactly the 12 skills claude plugin details lists, and 7 folders belonging to other plugins correctly left out.
- Reworded-request check: deepseek-v4-flash, 12 reworded requests plus 4 that need no skill. Offering every skill's name and description (160-character clip, 12 KiB budget) reached the right skill in 16 of 24 with memory on and 15 of 24 with memory off. None of the no-skill requests got one.
- Misses: the model answers directly or searches the workspace instead of opening the skill first. Weakest were PDF merge and extract, slides for investors, release notes and packaging a service. It needs one instruction line: when a listed skill fits, open it before doing anything else.
- Don't carry the word-match pre-pass. It attaches unrelated skills to any message containing "skill".
- Not handled yet: Claude Code's disable-model-invocation and user-invocable frontmatter (a natural fit for your turned-on subset), nested SKILL.md, CLAUDE_CONFIG_DIR, and the picker over --host reading the local disk.
We'll port the proof as tests onto feat/1277-skills once step 3 is in: the tmux end-to-end check, the reworded-request check and the real-install check, so the PR's acceptance is "these pass".

Sound right?

…s not one machine, and the dual lock is seen and taken back (#1392) The one-machine retry loop waits as its verdict says and never wraps to zero; only a request with no pool is one machine; the suite lock is seen by old checkouts and a dead directory lock is taken back under the flock. Full make check on Spark at e097553: EXIT 0.

Issue #1277's day-one list names .opencode/skills and .goose/skills beside the six folders discovery already read. They are appended after .gemini, so a name held in any of the first six still owns it and the later copy is kept in the result marked shadowed, in both the project and the home scope. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

AbirAbbas · 2026-09-30T18:00:32Z

Superseded by #1691: the same two roots re-cut onto current dev (this branch was stacked on #1410 and can't merge now that it's squashed), plus OpenCode's and Goose's real global folders under ~/.config and the manual page.

santoshkumarradha and others added 30 commits September 18, 2026 15:03

docs(manual): the roster's run rows are one line each, the run's row …

560f834

…ends in the dots

Merge c266: on the rail a plan row is one line and the run's row wear…

72f72a4

…s the dot row Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

hands-on: wave1 update (c266)

eeb8c60

hands-on: wave1 update

99c3a7f

hands-on: wave1 update (rail nesting)

2b6052e

docs: the 1204 entry says the rail draws the run's tree

e99fd79

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

test: reproduce woken planner ownership and terminal worker

908152f

fix: claim woken planners and stop terminal workers

aba6597

fix: restore ownership on every planner wake

9555b5d

docs: explain planner wake and repeated refusal

5bd4c74

hands-on: wave1 update (family stroke, dots)

c45fc53

tui3: a failure takes the dot row's end, one cell at least, and the r…

3ca7f70

…unning cell keeps the frontier Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

hands-on: wave1 update (dot cells)

b4bd365

Merge c269: a woken planner holds its task, and a worker whose task f…

dc91a73

…ailed is stopped Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

run: a refused claim never drops a wake, and a worker that wrote its …

9cc0af1

…own done is never cancelled Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

docs: the 1204 entry says a woken planner holds its task

cff326c

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

docs: the 1204 entry touches the engine too

a61547f

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

hands-on: wave1 update (woken planner holds its task)

7997f9e

docs: the run pane, one look says what this is and where it stands

0c275ff

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

test(session): specify stored run summaries

177ea1d

feat(session): store and refresh run summaries

0593e6f

feat(session): carry run now in landing digest

be16d7e

docs: explain run summary lines

4529f99

merge c271: the run summary, four lines on the worker model

9f17760

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

docs: the change entry says what the run summary invalidates

ea5113e

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

santoshkumarradha and others added 8 commits September 22, 2026 00:33

Merge origin/dev: sign in to Codex in the browser (#1336)

942eef6

# Conflicts: # internal/manual/chat/services.md

AbirAbbas force-pushed the feat/1277-skills branch from bf9fc6f to 69913c4 Compare September 23, 2026 16:16

santoshkumarradha added 3 commits September 23, 2026 12:24

tui3, chat, pool: task-page keys, notes, tab mark and unowned warms (#…

3a9c5e4

…1395) Twelve review findings on the task page, the tab mark, the pool judge sweep and the manual. Full make check on Spark at 635c029: EXIT 0.

session, config: the chat manages a live run in the side list's words…

b998c87

… and through every door (#1393) Nine review findings on the chat manager, the digest, config keys and the commit trailer. Full make check on Spark at 449e792: EXIT 0.

This was referenced Sep 23, 2026

santos/dev2 lands on dev: the worker harness as default, runs as work, skills from other harnesses, the chat as manager #1410

Merged

Skills from Claude Code plugins and Codex, with memory off, chosen by what they are for #1396

Merged

santoshkumarradha and others added 3 commits September 23, 2026 13:31

changes: the OpenCode and Goose skills folders entry

937397e

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

AbirAbbas force-pushed the feat/1277-skills branch from 69913c4 to 937397e Compare September 23, 2026 17:52

AbirAbbas changed the base branch from dev to santos/dev2 September 23, 2026 17:52

AbirAbbas changed the title ~~Skills: use SKILL.md folders from every harness, by name or on their own~~ skills: OpenCode's and Goose's folders are read too (on top of #1410) Sep 23, 2026

santoshkumarradha mentioned this pull request Sep 24, 2026

WIP review needed: skills tranche 1 (store, attachment, shelf, use_skill) #1327

Closed

Base automatically changed from santos/dev2 to dev September 25, 2026 02:20

santoshkumarradha added feature Work that adds a capability; developers break it into tasks area:tools The tool layer: built-in tools, registries, belts labels Sep 27, 2026

santoshkumarradha added this to the Reliable agent milestone Sep 27, 2026

AbirAbbas mentioned this pull request Sep 30, 2026

skills: OpenCode's and Goose's folders are read too #1691

Merged

AbirAbbas closed this Sep 30, 2026

Navigation Menu

skills: OpenCode's and Goose's folders are read too (on top of #1410) - #1409

skills: OpenCode's and Goose's folders are read too (on top of #1410)#1409
AbirAbbas wants to merge 852 commits into
devfrom
feat/1277-skills

AbirAbbas commented Sep 23, 2026 •

edited

Loading

Uh oh!

santoshkumarradha commented Sep 23, 2026

Uh oh!

santoshkumarradha commented Sep 23, 2026

Uh oh!

AbirAbbas commented Sep 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

AbirAbbas commented Sep 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What is in it now

Checked

Still to do here

Uh oh!

santoshkumarradha commented Sep 23, 2026

Uh oh!

santoshkumarradha commented Sep 23, 2026

Uh oh!

AbirAbbas commented Sep 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

AbirAbbas commented Sep 23, 2026 •

edited

Loading