Repository navigation
feat: AGENTS.md prompt support (closes #732) - #1161
Closed
Cedric Vidal (cedricvidal) wants to merge 14 commits into
Closed
Cedric Vidal (cedricvidal) wants to merge 14 commits into
Cedric Vidal (cedricvidal) wants to merge 14 commits into
Conversation
…rage Foundation for AGENTS.md prompt support (#732): typed task prompts ('task' | 'agents.md') with content-hash ids, size-based body storage (small bodies inline, large bodies uploaded to blob and referenced via contentBlobUrl), a prompt client/resolver, typed prompt-feature extraction, and request schema fields (agentsMd, agentsMdParentIds, agentsMdPromptId). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…ature routes Wire AGENTS.md support into the API (#732): create-request accepts agentsMd (raw text) and findOrCreates an 'agents.md'-typed prompt, storing agentsMdPromptId and agentsMdParentIds; task-prompt and prompt-feature routes become type-aware and expose a prompt content endpoint that resolves inline or blob-backed bodies. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The shared queue processor resolves the agentsMdPromptId body (downloading from blob storage when needed) and writes <workspace>/AGENTS.md once before the first turn. Fails the run loudly if the prompt id is set but cannot be resolved, rather than silently running the baseline (#732). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…ture commands Add CLI parity for AGENTS.md support (#732): 'run submit --agents-md <text|@file>' submits a run with an AGENTS.md prompt, and the task-prompt and prompt-feature commands gain a --type option to manage 'agents.md' prompts alongside task prompts. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add portal surfaces for AGENTS.md support (#732): an AGENTS.md field on the Submit Run page, AGENTS.md display on Run Detail, and type-aware task-prompt and prompt-feature views/pickers for 'agents.md' prompts. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…h API/CLI snapshots Document the AGENTS.md delivery flow and size-based prompt storage (PROMPT_INLINE_MAX_BYTES) in app-design.md and ENV_VARIABLES.md, and regenerate the OpenAPI and CLI command snapshots to reflect the agentsMd fields and --agents-md/--type options (#732). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
task-prompt.ts and prompt-feature.ts imported each other (task-prompt
needs PromptFeatureResultSchema; prompt-feature needs PromptTypeSchema),
forming a cycle. At ESM runtime the schemas barrel evaluates
prompt-feature first, so task-prompt's `z.array(PromptFeatureResultSchema)`
ran while the binding was still in the temporal dead zone, crashing the
API on boot with a ReferenceError. tsc and vitest tolerated the cycle, so
it only surfaced under the API's tsx runtime; under vitest the unresolved
binding also silently emitted an empty OpenAPI `features.items: {}`.
Extract PromptTypeSchema into its own neutral prompt-type.ts that both
modules depend on, removing the back-edge from prompt-feature to
task-prompt. task-prompt re-exports PromptTypeSchema so the barrel surface
and existing consumers are unchanged. Regenerate the OpenAPI snapshot,
which now correctly refs PromptFeatureResult.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…New Run The AGENTS.md input was always rendered as a prominent 6-row textarea even when unused. Make it discreet behind an "Add AGENTS.md" trigger that unveils a bordered card, mirroring the existing Codebase disclosure. Inside the card, add a TaskPromptPicker(type="agents.md") so an existing agents.md prompt can be looked up from the prompt library, just like the Task and gate prompts; selecting one fills the editor and the API dedups it by content on submit. Also restore AGENTS.md when prefilling from a recent run (resolve agentsMdPromptId to text) and surface an "AGENTS.md" chip in the settings summary. TaskPromptPicker previously imported the gates-only PromptType and indexed GATE_METADATA[type] for its placeholder, which has no agents.md entry. Switch to the broad PromptType from @/types and use promptTypeLabel(type) so the non-gate type renders safely; identical output for gate types. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Runs submitted with an AGENTS.md create an 'agents.md' task prompt, but the Prompt Library hid it: TaskPromptStore.getAll excludes NON_GATE_PROMPT_TYPES from the default (unfiltered) list. That guard is right for the report-template trigger picker, so instead of flipping the default this adds an opt-in 'includeNonGate' flag that the Prompt Library uses to list agents.md prompts alongside task/gate prompts. The library also gains an AGENTS.md type filter and create option; the CLI 'task-prompt list' gets a matching --include-non-gate flag. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The left-nav Library entry for /task-prompts pointed at the Prompt Library but was labeled 'Tasks', which read as a task list. Rename it to 'Prompts' to match the page title and the fact that it now manages task, gate, and AGENTS.md prompts. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The task-prompt list previously hid non-gate types (agents.md) from the default unfiltered list, requiring an opt-in flag to surface them. Remove that exclusion entirely so the Prompt Library, report-template pickers, API, and CLI always return every prompt type. Drops the includeNonGate option/flag/query param across the store, API route, portal client, and CLI, and regenerates the OpenAPI and CLI snapshots accordingly. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The detail page hard-coded "Task Prompt" / "Task Text", which is wrong for non-gate prompts such as AGENTS.md. Render the identity card title from the prompt type (e.g. "AGENTS.md Prompt", "Select Prompt") and rename the body card to the generic "Prompt Text". Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Adds focused unit tests for the portal changes in the AGENTS.md feature
that previously had no coverage:
- Layout: the Library nav item is labelled 'Prompts' and links to
/task-prompts (renamed from 'Tasks').
- TaskPromptDetail: the heading adapts to the prompt type
('AGENTS.md Prompt' / 'Select Prompt', legacy untyped falls back to
'Select Prompt') and the body card is the generic 'Prompt Text'.
- TaskPromptPicker: the AGENTS.md library lookup scopes the query to
type=agents.md, fills the field with the selected prompt text, and
shows an empty state when there are no matches.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Address review feedback: the GET /task-prompts/:id/content endpoint returns any prompt body by ID with no authorization check, consistent with the API's current no-auth/no-tenancy model. Since prompts are globally content-addressed and deduplicated via findOrCreate, document that this assumes a trusted-network, single-tenant deployment and must be revisited if auth/multi-tenancy lands. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Collaborator
Author
|
Original author: @github-actions[bot] Test Results (Node.js 22)test: Run #2036
🎉 All tests passed! |
Collaborator
Author
|
Original author: @github-actions[bot] Test Results (Node.js 22)test: Run #2043
🎉 All tests passed! |
Collaborator
Author
|
Original author: @github-actions[bot] Test Results (Node.js 22)test: Run #2044
🎉 All tests passed! |
Collaborator
Author
|
Original author: @github-actions[bot] Test Results (Node.js 22)test: Run #2045
🎉 All tests passed! |
Collaborator
Author
|
Original author: @github-actions[bot] Test Results (Node.js 22)test: Run #2046
🎉 All tests passed! |
Collaborator
Author
|
Original author: @github-actions[bot] Test Results (Node.js 22)test: Run #2051
🎉 All tests passed! |
Collaborator
Author
|
Original author: Wassim Chegham (@manekinekko) LGTM. |
Collaborator
Author
|
Original author: Cedric Vidal (@cedricvidal) Original review: COMMENTED |
Collaborator
Author
|
Original author: Wassim Chegham (@manekinekko) Worth a note here: this |
Collaborator
Author
|
Original author: Cedric Vidal (@cedricvidal) Good call. Added a comment on the route in 61bf19e4 spelling out the trusted-network / single-tenant assumption and flagging that the content-addressed dedup store is not tenant-safe, so this must be revisited if auth/multi-tenancy ever lands. Agreed it's not a regression given the API has no auth model today. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Original author: Cedric Vidal (@cedricvidal)
Why
Benchmark runs had no way to seed a project-level
AGENTS.mdinstruction file into the agent's workspace. Issue #732 asks for first-class support so a run can carry anAGENTS.mdthat the worker writes to the workspace root before the agent starts, and so those instruction files can be managed and feature-extracted like other prompts.This PR is the AGENTS.md slice extracted from the larger GEPA optimizer PR (#1031). It deliberately contains only the AGENTS.md feature and the size-based blob storage that underpins it.
experimentId, experiment grouping, the experiment migration, and thegepa-optimizerapp stay in #1031.Video walkthrough
agents-md-walkthrough.mp4
What changed
Typed task-prompt model +
agents.mdtype (shared). Task prompts now carry atype(PromptType = GateId | "agents.md"). AGENTS.md content is stored as anagents.md-typed task prompt, deduplicated by content viafindOrCreate, so reusing the same body reuses the same prompt. Prompt-feature extraction is type-scoped (task prompts get task features; AGENTS.md prompts getagents.mdfeatures).Size-based blob storage (shared). Prompt bodies above
PROMPT_INLINE_MAX_BYTESare offloaded to blob storage with acontentBlobUrl, and resolved transparently through a/contentendpoint. This is the foundation for large AGENTS.md bodies; small bodies stay inline as before.Worker delivery. Before a run starts, the worker writes the resolved AGENTS.md to
<workspace>/AGENTS.md.Surfaces. API request handling (
agentsMdresolves to anagents.mdprompt) plus typed task-prompt and prompt-feature routes; CLI--agents-mdrun flag and type-awaretask-prompt/prompt-featurecommands; portal support to submit an AGENTS.md with a run, view it on a run, and manageagents.mdprompts.New Run UX. The AGENTS.md field is now discreet: hidden behind an "Add AGENTS.md" trigger that unveils a card (mirroring the existing Codebase disclosure), and it can look up an existing
agents.mdprompt from the library via the same picker used for task and gate prompts.Worth a careful look
fix(shared): break task-prompt/prompt-feature schema import cycle(ccfce549).task-prompt.tsandprompt-feature.tsimported each other, which threw a temporal-dead-zoneReferenceErrorunder the API'stsxruntime on boot (tsc and vitest tolerated the cycle, so it passed CI but crashed the running API). The fix extractsPromptTypeSchemainto a neutralprompt-type.tsthat both modules depend on, removing the back-edge. The cycle had also silently emitted an emptyfeatures.items: {}in the OpenAPI spec; it now correctly refsPromptFeatureResult(regenerated snapshot).openapi-snapshot,cli-commands) were regenerated, not hand-edited.Verification
pnpm -r buildclean.apps+packages(excluding snapshots): noexperimentId/gepa/GEPA.Closes #732