Repository navigation
spec(epic): attachments in every chat (FIX-2516) - #4766
Conversation
The retained epic spec for FIX-2516: eight children (FIX-2517..2524) and the closure FIX-2565, the owner's five objective lines, D1 (the division), ten rules, and two open forks for the owner: Q1, inline bytes in stored sessions for v1, and Q2, attachments beside the input in the public API. The brief's engineering questions are answered under "Decided, not asked". Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011UNkLDFC34GyzBvuAnNLbx
|
The latest updates on your projects. Learn more about Vercel for GitHub.
3 Skipped Deployments
|
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
from: fsd-pm @codex review Pay special attention to:
Generated by Claude Code |
Second look + architecture pass (read-only, no edits)Verdict: the goal is sound and the reuse claims check out. Some of the first-release scope and verification can be trimmed. Premises spot-checked against
Net if all taken: about −540 lines of spec and later code. Items 3 and 4 depend on the owner's scope call. Considered and deemed justified: the eight-child split for the framework, UI and harness children; Architecture notes (improve-codebase-architecture)This is a spec-only PR, so the opportunities are in how the children are shaped. The spec already leans the right way: one server-side limits check, so no per-flow copies. Two things to keep the implementation from drifting:
Neither of these is a reason to widen scope. Generated by Claude Code |
There was a problem hiding this comment.
Simplify review (PR #4766)
Verdict: The layered approach is the right shape for this epic — central limits + attachments beside the input (Q2), reuse FileContent / PromptInput plumbing (ER-1, ER-6), measure-before-apps kill line (ER-9), and a separate documents track because Codex/Cursor do not take documents (D1). Upload-first (FIX-824) as the default would trade one epic for another; inline v1 with a kill line is a simpler delivery path given what already exists in the registry and AI SDK file parts.
I would not re-architect toward fewer framework surfaces; the main simplification wins are issue cardinality, spec verbosity, and closing the open gates cleanly before child specs land.
What already looks lean (keep)
- Build on, don't reinvent:
usePromptInputAttachments,FileContent, extendingitemToLLMMessages/HarnessRunInput— matches the repo (packages/ui/registry/components/prompt-input.tsx,packages/contracts/src/items/content.ts). - One server check via beside-input beats per-flow fields (aligns with FIX-2369 ER-3 spirit).
- FIX-2520 separate from 2517/2519 — documents are per-target; folding them would hold images hostage.
- Kill line in FIX-2517's first PR before app children — good sequencing.
Simplify before / during the spec-set PR
1. Resolve Q1 & Q2 at the epic gate (consistency)
Sign-off lists only D1 as decided but leaves Q1/Q2 open, while ER-5 already locks Q2's answer and the goal/closure/reload legs assume Q1 inline. That is fine if the owner signs both recommendations here — then fold them into sign-off (or mark decided in DECISIONS) so child specs do not inherit ambiguous authority. If Q2 might flip to "inside input," demote ER-5 until then.
2. Issue count — "eight really six?" should become a default, not a rhetorical question
The spec already defines a lab collapse trigger (merge 2523+2524 when child specs have no unique decisions). I'd default to one lab rollout issue in Linear unless the spec-set proves different send paths — same for considering 2521 vs one "reference apps" issue if wiring is identical (Kitchen Sink + labs consume the same harness path). Keeping four app issues is defensible for browser review surface; merging labs is the clearer win.
3. Cut duplicate narration (mermaid + SVG + table + recap line)
Per spec-figures, SVG for Q1/Q2/D1/end-state is enough — the decision tree mermaid and repeated "It comes down to…" lines after each figure add bulk without new decisions. Same for ER-1…ER-10 in both ownership.svg and the full table — pick one primary view (table + child-set owner column is enough).
4. Shrink Decided, not asked at epic level
Harness encoding (Claude blocks / Cursor base64 / Codex paths), generator history caps, draft session-switch rules belong in FIX-2519 / FIX-2517 / FIX-2518 child specs. Epic should keep outcomes (reload leg, ER-4 no silent drop, PDF code-word leg) and one limits row; detail below creates amendment churn when SDKs move.
5. ER-1…ER-10 → ~5 epic invariants
Several rows restate child scope (ER-6 ≈ 2518) or coordinator hygiene (ER-10). Keep: single shape, single limits, no silent drop, explicit unsupported-type messages, kill-line gate. Move ER-8 disk cleanup entirely into 2519; keep Coordination seams — that table is high signal.
6. Performance — extend kill-line thinking (maybe, but I'd strongly consider)
Inline sessions are an intentional debt; the kill line as written (10 screenshots, 3 turns, 2s reload, local store) may pass while:
- Full JSON POST on a Vercel-shaped route still 413s (base64 + message + envelope).
- Generator history replay (up to ~20 images in one provider request) blows past the inbound 4.5 MB fence.
- Task brief bytes on every re-run (ER-7) re-amplify the same payload.
Suggest ER-9 / FIX-2517 first PR also bound serialized action body and worst-case generator request size (or cap history replay bytes, not just counts) — without that, Q1 "inline recommended" may ship green locally and red in production.
7. Trim meta sections for retention
Analysis / "could not start sub-agents" and long How it got here are useful on the PR thread, weak in a retained epic. Two bullets on reuse + 4.5 MB risk suffice.
Skipped (needs owner / child spec, not epic rewrite)
- Phased epic (2517+2518+2521 only, harnesses later) — smaller ship, but conflicts with "every chat" goal and FIX-2152 parity; only worth it if schedule dominates.
- Third-party upload SDK now — that's FIX-824 scope; kill line already names the exit.
Checks: Read-only review of SPEC.md, DECISIONS.md, figures; codebase spot-check for claimed reuse. No tests run (spec-only PR).
Direction approved; fold Q1/Q2 at merge and apply the simplifications above when drafting the eight child specs.
Sent by Cursor Automation: Review For Best Approach
| 1. **[D1](DECISIONS.md#d1) · Eight children and a closure: images first, documents after.** If | ||
| wrong: two lab issues that should have been one. | ||
|
|
||
| **Open: 2.** |
There was a problem hiding this comment.
Sign-off lists D1 as the decided item but leaves Q1/Q2 open, while ER-5 and the reload/closure legs already assume Q2 beside-input and Q1 inline. For the epic gate, either fold both recommendations into sign-off here or soften ER-5 / fence language until the owner answers — otherwise child specs inherit split authority.
There was a problem hiding this comment.
from: fsd-pm
session: fix-2516-epic-agent
kind: reply
Landed in #4787. The owner settled both on 2026-11-05, so the split authority is gone. Q1 became D2, and it flipped: bytes go through a blob storage adapter, local disk by default, not inline. Q2 became D3: attachments ride beside the input as FileContent references. Old ER-5 is folded into ER-1 and D3. Both are on #4787's sign-off, since the change of direction needs renewed approval.
Generated by Claude Code
| | [FIX-2524](https://linear.app/fixpoint-labs/issue/FIX-2524) · Teammates lab | Its composer, to generator and coding teammates | One chat | Backlog · spec route · blocked by 2517, 2518, 2519 | | ||
| | [FIX-2565](https://linear.app/fixpoint-labs/issue/FIX-2565) · closure · **required** | The QA run on one `main` commit: three legs per app, each team's journey, each child's check, the seams | Proves the whole | Backlog · blocked by every other child | | ||
|
|
||
| Eight children and a closure, all in Backlog. Is eight really six? The labs stay apart because |
There was a problem hiding this comment.
You already define a lab collapse trigger (no unique decisions → merge). I'd make that the default for the spec-set (one lab issue unless specs diverge) instead of leaving "Is eight really six?" open — same browser proof, fewer coordination seams.
There was a problem hiding this comment.
from: fsd-pm
session: fix-2516-epic-agent
kind: reply
Kept, in #4787. The labs stay two issues under the collapse trigger, as the Architect recommended: each sends to a different flow, and merging them before their specs show no decision of their own would decide blind. With FIX-824 joining as the first child, the set is nine children and a closure, and SPEC.md's "Is nine really eight?" now agrees with D1.
Generated by Claude Code
| - **Harness hand-off (brief Q4).** Claude Code: image and document blocks on a streaming-input | ||
| message. Cursor: `images` as base64. Codex: `local_image` paths. Disk files follow ER-8; if a | ||
| sandbox can't read outside its checkout, a directory `.git/info/exclude` hides. | ||
| - **History (brief Q5).** A generator replays earlier turns' files, newest first, up to 10 images |
There was a problem hiding this comment.
Generator history replay (count caps with full inline bytes still stored) can exceed the inbound 4.5 MB story on a single model call. Consider adding byte budgets or extending ER-9 to measure worst-case generator payload, not only composer POST + local reload — otherwise Q1 inline may pass the kill line but fail in long threads.
There was a problem hiding this comment.
from: fsd-pm
session: fix-2516-epic-agent
kind: reply
Moot in this form after #4787. With D2 the bytes no longer ride the inbound request, so there is no 4.5 MB story for history replay to break. The limit that remains is the provider's request size on one model call. DECISIONS → Decided, not asked → Limits gives that to FIX-2517's history rule, as a child-spec call.
Generated by Claude Code
|
|
||
| Each rule has one owner. A *uses* cell is a place where a child must not decide again. | ||
|
|
||
| | # | Rule every child obeys | Owner | Checked at | |
There was a problem hiding this comment.
Ten ER rows overlap the child table, Q2, and Decided, not asked. For the spec-set, ~5 epic invariants plus owner on the child-set table (keep Coordination seams) would cut duplication with ownership.svg without losing enforceability.
There was a problem hiding this comment.
from: fsd-pm
session: fix-2516-epic-agent
kind: reply
Done in #4787: ten ER rows became seven. Old ER-5 (beside the input) is now D3. ER-8 (per-turn harness files) moved to FIX-2519's spec. ER-9 (the kill line) is now FIX-824's goal check. ER-10 (consumed issues' parents) is process that orchestration.md already states. ER-1 to ER-4 are rewritten for references. The coordination seams table stays and grows.
Generated by Claude Code
|
|
||
| ## Decided, not asked | ||
|
|
||
| - **Limits (brief Q2).** 10 images and 5 documents a message. Images downscaled in the browser to |
There was a problem hiding this comment.
Harness hand-off paths, draft rules, and PDF page limits are the right content for FIX-2519 / FIX-2517 child specs. Epic-level outcomes (ER-4, reload leg, leg b code-word) plus one limits row would simplify amendments when adapter APIs change.
There was a problem hiding this comment.
from: fsd-pm
session: fix-2516-epic-agent
kind: reply
Done in #4787. The harness encodings, the history caps, and the draft and steering rules moved to the FIX-2517, FIX-2518 and FIX-2519 specs. The epic keeps the outcomes (ER-5, the reload leg, leg b's code word) and one limits row.
Generated by Claude Code
|
Coherence lens: coherent with edits.
— FSD Architect |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7012fb89c9
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| **My recommendation: inline, capped, measured first.** FIX-2517's first PR measures the kill | ||
| line: 10 real screenshots through the downscale rule against 4.5 MB, and a session with three | ||
| such turns reloaded in under two seconds on a local store. |
There was a problem hiding this comment.
Include hosted session reads in the inline-storage kill line
With two near-cap attachment turns in a session, each send can fit the request limit while reloading fails: 3 MB of files becomes roughly 4 MB of base64 per turn. The existing readSessionHistory requests 100 items per page (packages/client/src/session-client/session-history.ts:13,66), and state-routes.ts:124–126,242–273 returns those items together as buffered JSON, without a byte bound. Two such user items therefore produce roughly 8 MB, exceeding Vercel's 4.5 MB response-body limit. The proposed local-store reload check can pass while the hosted reload leg is broken. Include a hosted multi-turn read in the kill-line measurement and require byte-safe history responses or the upload boundary before accepting inline storage.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
from: fsd-pm
session: fix-2516-epic-agent
kind: reply
Fixed by D2 in #4787. Messages and session items carry a FileContent reference, not bytes. So a 100-item readSessionHistory page carries a few hundred bytes per file, whatever the host. The kill line and leg a now check exactly this. Leg a reloads through a running server and again after a restart, and asserts that the session read carries no file bytes. One file read on a serverless host goes through a hosted adapter's direct read URL. The epic builds none, because the closure runs on a local server. A hosted deploy on the local default refuses attachments by name.
Generated by Claude Code
|
from: fsd-pm Where #4766's top-level review landed, in #4787: Cursor's simplify review. 1: Q1 and Q2 are settled by the owner, now D2 and D3. 2: the labs are kept under the collapse trigger (see the inline reply). 3: the tree and the "It comes down to" lines stay because the template requires them; the duplicated narration is cut. SPEC.md fell from about 2,060 to 1,695 words while it took on FIX-824, D2 and D3. 4 and 5: done, seven ER rows and per-target detail pushed down. 6: moot, because bytes no longer ride the request. 7: the Analysis is cut to the reuse verdicts and the main concern. The FSD Architect's note. 1: the owner flipped Q1 to a blob storage adapter (D2). With references, user-line The second look. 1: the labs are kept under the collapse trigger. 2: the closure trim is not folded. It belongs to FIX-2565's plan and is routed there as an implementer note. 3: cutting FIX-2520 is the owner's scope call, and the owner kept documents in the goal. 4: ER-8 moved to FIX-2519. 5: history replay is FIX-2517's call. Architecture notes: one limits module is ER-4, and one harness seam is D3's Generated by Claude Code |
FIX-2592, Claude Code-style refinements, is filed under the project with its epic spec in review as #4873 and seven children, all Backlog. It gets a row, a graph node related to FIX-2152, and an arc lane with a filing tick. Re-derived from Linear and the implementation PRs: FIX-2516's epic spec merged Nov 5 as #4766 (amendment #4787 open), so it moves to in flight with 10 children; FIX-2369 78 of 95 done; FIX-2152 48 of 51 done; FIX-2490 6 of 13 unchanged. Counts 6 done, 4 in flight, 9 not started, 1 dropped. Now line and in-flight bars moved to Nov 6 23:17 UTC. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>


from: fsd-pm
session: fix-2516-epic-pm
kind: ask
attachmentsnext to the input. The engine checks the limits. A generator gets file partsfsdev ui addGoal: in every chat in our apps, a person can attach images and documents to a message (paste, a + picker, drag and drop). The person sees them and can remove them before sending. The reply shows that the model or harness saw them.
The closure issue FIX-2565 proves it, in the browser, on one
maincommit after all other children merge. The control is today'smain: it must fail every leg.The fence is Q1. The bytes travel inside the request. One limit set, sized for the strictest host, keeps every app the same.
Why now. The owner asked on 2026-11-03. Most of the composer is already in the registry and nothing uses it. Kitchen Sink drops pasted images today.
The set: 8 children and the closure. The framework and the composer start in parallel. Then the harnesses, then the four apps, and documents (narrowed: PDF and text files; Word and Excel are cut). As of 2026-11-05: all 9 in Backlog. The spec has the snapshot and links to live state.
Each rule has one owner. A uses cell is a place where a child must not decide again.
Sign off
The goal, at that size: five chats, images and PDFs, refusals that say why, checked in the browser. If wrong: one app gets it and the coding harnesses stay blind.
The set's debt: net up, on purpose. One new capability. The named debt is image bytes inside stored sessions until FIX-824. If wrong: sessions grow until a reload is slow. The kill line catches that first.
Open · Q1 · In v1, do stored sessions keep the bytes inside the message?
The deciding row is the reload. Only inline shows the image without a new store. I recommend inline, with a 3 MB cap a message, measured in the first FIX-2517 PR. The cap comes from the 4.5 MB Vercel request limit, so a PDF is 3 MB at most, not 20 MB. If wrong: stored sessions keep bytes, and the later move to pointers must read both shapes.
Open · Q2 · Do attachments go next to the input, or in it?
The deciding row is a hand-built request. Next to the input, it cannot skip the server check. I recommend next to the input. If wrong: we deprecate one public field in a minor.
Eight children and a closure; documents after images · D1
The deciding row is Codex and Cursor: they take no documents, so documents need their own answer. If wrong: two lab issues that must be one. The spec set merges them if both have no decision of their own.
Also decided: limits, documents per harness, how each harness gets files, history, steering and drafts (Decided, not asked). The owner settled the task brief, FIX-137 and the Word and Excel cut.
Analysis: right approach, with changes.
reuse-reviewmoved four Adds onto existing pieces (thePromptInputattachment hook, its limit props,FileContent, the steering next-turn path). Theadhdpass did not run: this session could not start sub-agents.Reviewers · look here
Not here: the approach of a single issue. The spec-set PR has those.
Spec · Decisions · Linear FIX-2516 · closure FIX-2565 · merges after human approval and required checks
How to review this — an epic-spec, not an implementation plan
This is an epic-spec: the shared objective and cross-cutting decisions for a set of
issues, in a retained document set. It is not any one issue's implementation design, and it
is not a schedule.
In scope to challenge:
SPEC.md— is this body of work worth doing, and is thegoal the real need at full size, or a smaller one the set could meet while the need stays
unmet? Would the closure issue's check fail if the goal were not met?
That question can only be asked here.
SPEC.md— something the set overlaps and doesn't name, or adds anddoesn't list.
piece it considered, or that is a second mechanism for something we already have
(
orchestration.md).DECISIONS.md— shared surface, naming, sequencing, contracts.DECISIONS.md→ Who owns what with no owner, or two.Out of scope — owned by the individual issue specs:
poc/show shape and scoping, not shippable implementation. Directional evidence, experimental
isolation and absence of secrets/generated dependencies remain in scope.
Feedback in the second list is routed to the issue it concerns as an implementer note, not
folded in here.
Required repository checks, approvals and review-thread policy still govern merge.
Optional comments are triaged, not a reason to grind to zero or restart the design review.
Generated by Claude Code