Process Inbox

Process Inbox

When running in Grok Build, read .grok/ai-workflow-adapter.md completely before any other host-specific delegation rule and apply its substitutions.

Turn the human-written ignored inbox into tracked planning artifacts. Route and plan only: do not edit Telegram source, build, test, claim, or implement tasks.

Workspace

Run from a Telegram Desktop checkout. Use the bundled helper with an available Python 3 interpreter:

python3 .agents/skills/process-inbox/scripts/workspace.py inbox-ensure
python3 .agents/skills/process-inbox/scripts/workspace.py prepare

Use python or py -3 when that is the host's Python 3 command. The helper:

• reads Telegram/build/ai-machine-tag;
• combines it with the checkout folder, for example macbook-twork;
• locates the sibling ai-tdesktop and ai-tdesktop-worktrees directories, with AI_TDESKTOP_ROOT and AI_TDESKTOP_WORKTREES_ROOT as overrides;
• ensures the isolated inbox/<checkout-tag> linked worktree exists without creating or inspecting the task execution worktree;
• snapshots the ignored inbox before planning and prints JSON paths.

prepare resumes the one active .processing-* snapshot when present. Save its transaction, digest, ai_main, inbox_worktree, inbox_branch, and checkout_tag values. Read raw input only from the transaction snapshot, not from the live inbox.

The task execution slot_worktree may be dirty or have unpublished task work. Ignore it completely while processing the inbox: do not read planning state from it, write to it, stage it, commit it, rebase it, or publish it.

For a new transaction, prepare requires clean ai_main and inbox_worktree state with no unpublished inbox commits. It fetches origin when configured, fast-forwards local master, and fast-forwards the inbox branch before taking the snapshot. If the inbox is empty, the machine tag is invalid, or either inbox publication worktree is unsafe, stop without changing or clearing the inbox. Never force-push shared AI history.

Route and plan

Read these before planning:

• source checkout AGENTS.md;
• ai_main/AGENTS.md;
• existing <inbox_worktree>/projects/*/project.md, including projects/archive/, and relevant task states from <inbox_worktree>;
• the transaction's inbox.md and every file it references.

Some retained task directories have superseded.yaml instead of state.yaml. They are durable aliases created by queue consolidation, not missing or reusable paths. Follow superseded_by chains to their live task when deduplicating, resolving prior receipt references, or checking whether a same-digest result still exists. New dependencies and project links must name the final live task, never an alias. A dated slug occupied by an alias still counts as a collision. Some retired oversized tasks instead have split.yaml, whose split_into list resolves to several live successors. Treat the retained directory as history, deduplicate against all live successors, and never put the split id in a new dependency or project link. A split path is also permanently occupied.

Use one disposable leaf planner when the harness supports delegation; instruct it not to delegate. Otherwise perform the same work locally. The planner may write a proposed routing file inside the ignored transaction, but only the orchestrator writes tracked AI state.

Treat natural-language hints as evidence, not required syntax. Segment the inbox into requests, then decide for each request whether to:

• create a standalone task with no project;
• add one or more tasks to an existing project;
• create a new project when durable shared context is useful.

Never create a task whose work is to move an existing source commit between branches: no backport, forward-port, cherry-pick, rebase, merge, branch sync, or equivalent integration task. Branch placement is human release/history coordination, not product work for the autonomous queue. When a request only asks for that operation, record a receipt-only disposition naming the existing source task or commit description and the requested target branch, then leave the operation to the human. When new product work requires code shipped by an earlier task, route only the product work and express the source task as a dependency; do not create a companion task to bring that dependency onto a branch.

Bias project assignment toward continuity. When a request follows from an existing task, begin with that task's project and keep it unless independence is affirmatively established. A task belongs to the existing project when it builds on project code or behavior, requires source changes shipped by project tasks, assumes the project's branch or accumulated context, or must be ordered after project work. Touching shared infrastructure or an additional non-project consumer does not by itself make the task standalone: projects record feature and code lineage, not exclusive ownership of every touched file.

Route a derived request to another project or to project: null only when it remains coherent, implementable, and independently testable in a checkout where the originating project's changes are absent or reverted. Record that concrete independence evidence in the receipt; "cross-cutting", "cleanup", or "broader than the source task" is not enough. For a request without a source task, prefer an existing project whenever its code, plan, or prior tasks supply essential context, and use a standalone task only when no project does.

Do not create generic holding projects such as fixes. A release batch of unrelated regressions normally becomes standalone tasks or tasks in existing domain projects. Group requests into one task only when they form one cohesive, independently testable behavior that can be implemented, reviewed, and tested as one normal pass.

Split at product boundaries, not arbitrary file or line-count boundaries. A request needs multiple tasks when two or more parts have their own useful outcome and acceptance oracle, when a later part can consume an earlier part as a stable dependency, or when the parts require materially different context, failure analysis, or evidence setup. New network, persistence, concurrency or ownership machinery plus application lifecycle/UI integration are especially strong split signals when each can be exercised independently. Shared project context, overlapping files, or one eventual feature does not by itself justify paying one review and test loop over their combined implementation.

Do not over-split inseparable changes: keep a small API and its only caller together when neither has a meaningful standalone result, and keep one atomic behavior together when separating it would leave an unbuildable or untestable intermediate state. For every proposed task, state the one shipped boundary it owns and the direct evidence that can approve it without first implementing a sibling. If that sentence needs several independent outcomes or several unrelated instruments, split again.

This is the first scope gate, not an irrevocable ruling. Inbox planning uses the request plus light source inspection and deliberately does not construct the implementation plan. The later independent perform-task assessment sees exact files, APIs, phases, ownership boundaries, and evidence design; it may veto the single-task shape when that richer proof exposes independently shippable and testable boundaries. That veto does not mean task sizing is based on elapsed time or diff length, and it does not authorize the performer to mutate the queue itself. The performer publishes split-required; the checkout scheduler then launches a dedicated deep split transaction that creates replacements, rewrites dependencies and project links, and retires the source with a durable multi-target split record.

Project slugs are unique across projects/ and projects/archive/. When a request belongs to an archived project, restore it before routing to it:

python3 .agents/skills/process-inbox/scripts/workspace.py inbox-unarchive \
  --project <slug>

The helper moves the project back to projects/<slug>, rewrites its relative links, and leaves the restored files staged for this transaction's commit. Never point a task at a path under projects/archive/.

Briefly inspect Telegram source when needed to understand scope and testable seams. Do not plan implementation internals and do not modify the source tree.

Assign task paths

Use the processing date and a concise imperative kebab-case slug:

tasks/YYYY/MM/DD/<task-slug>/

The task identifier is the path below tasks/, for example:

2026/07/18/fix-community-forward

Never ask the human to choose or remember it. Consult existing directories and append -2, -3, and so on to resolve a same-day collision. Dependencies may name only task identifiers created earlier in the same routing result or existing tasks.

Write tracked artifacts

Write tracked planning artifacts only inside the checkout-specific inbox_worktree.

For every task, create task.md:

# <imperative title>

<self-contained request and relevant constraints>

## Acceptance

- <specific observable result proving the behavior>

## Inputs

- [<descriptive label>](input/<file>)

Omit Inputs when none are used. For visual work, include the design basis and the exact visual/layout evidence expected. Copy every pertinent supplied file into input/; never reference the ignored inbox or its backup from a task.

Keep acceptance criteria to what actually proves the requested behavior. Never write a test-data integrity criterion: the live TelegramForcePortable folder is a disposable copy, so a task must not ask a performer to hash, back up, compare, or restore the account, to verify any setting or folder is unchanged, or to leave the account as it was found. A run may leave templates, theme, wallpaper, window geometry and interface scale mutated. State needs restoring only where a later measurement in the same run depends on it, never as an end-of-run obligation. The test-loop's SETUP owns the folder — it requires the golden test_TelegramForcePortable, reuses a live folder carrying the testing marker, and otherwise prepares a fresh copy — and its test-run command already launches with -testagent -noupdate, so no task needs to require either flag. Every such criterion costs a performer real time and proves nothing about the product.

Never write an acceptance criterion that can only be satisfied by adding debug machinery to production code — #ifdef _DEBUG blocks, debug-only types, observation structs, counters, or hooks in product translation units. Observability belongs to the disposable overlay and the permanent Telegram/SourceFiles/test/ helpers (see "Debug-Only Code" in the source checkout's AGENTS.md). A request whose proof seems to demand production instrumentation is misdesigned: route the product behavior, and let the performer's harness own how it is observed.

Create state.yaml in this exact field order:

status: todo
type: implement
created: YYYY-MM-DD
project: null
depends_on: []
claimed_by: null
claimed_at: null
claim_order: null
lease_until: null
phase: null
carried_from: null
inbox_receipt: receipts/YYYY/MM/DD/<receipt>.md

Do not write a model field. It records which model finished the task, so only finish writes it, at the canonical Approve, Block, or Split-required boundary; a task carrying one before it is claimed is malformed.

Use a project slug instead of null when routed to a project. Use a YAML list of task identifiers for dependencies. Dependencies record code lineage as well as readiness: keep an approved source task in depends_on when the new task's implementation assumes its shipped changes. State that prerequisite in task.md. This dependency is sufficient; never add a separate backport, cherry-pick, rebase, merge, or branch-sync task to make it reachable. Inbox processing never reserves work: new tasks always remain status: todo with claimed_by, claimed_at, and claim_order set to null. The checkout tag belongs in the receipt only.

Every new task uses type: implement. Do not predict a cheap or expensive execution profile while routing: the performer selects review specialists and evidence instruments after inspecting the real code and risks. Keep acceptance criteria outcome-focused. Name a required build, probe, app run, interaction, or screenshot only when that instrument is itself part of the requested result or no other instrument could decide the claim from the facts already known to the planner.

For a new project, create projects/<slug>/project.md with a concise durable scope and projects/<slug>/tasks.md with task links. For an existing project, append every newly assigned task link, including inherited follow-ups. Project indexes do not store live status.

Create one tracked Markdown receipt under receipts/YYYY/MM/DD/. Include:

• local processing time, checkout tag, and inbox digest;
• every inbox request mapped to friendly task titles and identifiers;
• every supplied file mapped to its copied task input, or explicitly unused;
• created projects and updated projects;
• deduplication decisions.

Before writing, search receipts for the same digest. If it was already fully processed and every referenced task either has live state, has a durable alias chain reaching live state, or has a durable split record whose successors all resolve to live state, create nothing and reuse that receipt for finalization.

Validate and publish

Before committing, verify:

• every request and supplied file is accounted for;
• every new task has task.md, valid state.yaml, and a falsifiable acceptance result;
• every task link, dependency, and copied input exists;
• no task or project reference points into projects/archive/;
• no raw inbox path, .local/, browser profile, portable account, credential, complete run directory, or complete build log is tracked;
• no Telegram or AI commit hash is copied into a task, project, or receipt;
• only expected tasks/, projects/, and receipts/ paths changed;
• tracked text uses the checkout's native convention (LF on Unix/WSL, CRLF on native Windows) without a BOM or mixed line endings.

Publish through the helper, passing every generated or restored task, project, and receipt path explicitly:

python3 .agents/skills/process-inbox/scripts/workspace.py inbox-publish \
  --transaction <transaction-path> \
  --receipt <receipts/YYYY/MM/DD/name.md> \
  --path <tasks/YYYY/MM/DD/task-slug> \
  --path <projects/project-slug> \
  --path <receipts/YYYY/MM/DD/name.md>

Do not run broad staging commands yourself. The helper rejects changes outside the explicit paths, stages those paths, commits them on the inbox branch with Process inbox for <checkout-tag>, fetches and rebases onto current master, pushes HEAD:master when an origin exists, and fast-forwards local master. It can resume publication when the inbox commit already exists. Ordinary non-fast-forward races are retried. If a semantic rebase conflict, unsafe inbox worktree, or remote outage occurs, do not clear the inbox; leave the active transaction and inbox commit recoverable and report the exact state. Never cherry-pick or force-push, and never involve the dirty task slot. When a project was restored, pass both projects/<slug> and its removed projects/archive/<slug> path.

After master contains the generated commit and any configured push succeeded, finalize using the receipt path relative to ai_main:

python3 .agents/skills/process-inbox/scripts/workspace.py finalize \
  --transaction <transaction-path> \
  --receipt <receipts/YYYY/MM/DD/name.md>

The helper accepts only a receipt below receipts/, verifies that exact receipt and digest are present in local master HEAD, checks the live inbox digest, preserves the raw snapshot under ignored inbox/backup/, and empties inbox.md only when the input remained unchanged. If the human edited the inbox during planning, it preserves those edits and reports cleared: false.

If planning fails before any tracked changes or commits exist, preserve the live inbox and close the snapshot with:

python3 .agents/skills/process-inbox/scripts/workspace.py abort \
  --transaction <transaction-path>

Do not abort after a generated inbox commit exists; retain the transaction so publication can be resumed.

Report

Return a compact summary with friendly task and project titles, the AI master publication status, the local backup path, whether the inbox was cleared, and any unused input. Do not report a commit hash or start implementation automatically.