Phase 6 of the AWS ECS/Fargate host (spec docs/superpowers/specs/2026-11-13-aws-ecs-host-design.md, plan docs/superpowers/plans/2026-11-15-aws-phase6.md).
What it adds
EFS storage mode — /workspace on a second access point of the user's own file system; repos and builds survive a stop.
Copy-at-connect — credentials are copied into the workspace after each connect instead of stored in Secrets Manager; ~/.ssh stays off EFS; switching an existing project recreates its secret so no old version keeps credentials.
Setup scripts — edited in the Config tab, run as claude after the clone on every start (30-min limit); progress and failures in the Overview card. Start finishes at "Connecting"; clone/setup are followed in the background and Stop cancels them.
Environment lifetime cap — max_workspace_hours on the environment (module v2, so existing environments show Update available); enforced inside the workspace (advisory), exit code 76, shown in the project's settings.
Port mappings on AWS — loopback-only forwards through the agent tunnel; the sidecar's port is refused.
LLM gateway sidecar — optional LiteLLM container in the task with a per-start master key; config via os.environ/ references so no key sits in the task definition.
Cost estimate — per workspace from the AWS Pricing API (cached 24 h; adds pricing:GetProducts to workspace-user-tasks.json).
WIP branches — one per run per folder, updated in place; unchanged trees skipped.
Live (us-west-2): lifetime value reaches the task, EFS persistence across restart, setup script, port mapping, sidecar health with no keys in logs, copy-at-connect (no credentials in the secret, CA installed), cost estimate. < $0.05, account left clean.
Not exercised live: the lifetime cap firing, a v1→v2 environment update, the secret recreation on switch.
Phase 6 of the AWS ECS/Fargate host (spec `docs/superpowers/specs/2026-11-13-aws-ecs-host-design.md`, plan `docs/superpowers/plans/2026-11-15-aws-phase6.md`).
## What it adds
- **EFS storage mode** — `/workspace` on a second access point of the user's own file system; repos and builds survive a stop.
- **Copy-at-connect** — credentials are copied into the workspace after each connect instead of stored in Secrets Manager; `~/.ssh` stays off EFS; switching an existing project recreates its secret so no old version keeps credentials.
- **Setup scripts** — edited in the Config tab, run as `claude` after the clone on every start (30-min limit); progress and failures in the Overview card. Start finishes at "Connecting"; clone/setup are followed in the background and Stop cancels them.
- **Environment lifetime cap** — `max_workspace_hours` on the environment (module v2, so existing environments show *Update available*); enforced inside the workspace (advisory), exit code 76, shown in the project's settings.
- **Port mappings on AWS** — loopback-only forwards through the agent tunnel; the sidecar's port is refused.
- **LLM gateway sidecar** — optional LiteLLM container in the task with a per-start master key; config via `os.environ/` references so no key sits in the task definition.
- **Cost estimate** — per workspace from the AWS Pricing API (cached 24 h; adds `pricing:GetProducts` to `workspace-user-tasks.json`).
- **WIP branches** — one per run per folder, updated in place; unchanged trees skipped.
## Testing
- `cargo test --locked`: 1,506 passed · `vitest`: 1,322 passed · build clean · image builds.
- Live (us-west-2): lifetime value reaches the task, EFS persistence across restart, setup script, port mapping, sidecar health with no keys in logs, copy-at-connect (no credentials in the secret, CA installed), cost estimate. < $0.05, account left clean.
- Not exercised live: the lifetime cap firing, a v1→v2 environment update, the secret recreation on switch.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
EFS storage is Hybrid plus /workspace on a second access point of the
user's own file system, rooted at /triple-c-workspaces/<owner_key>/<project>
(a sibling of the home roots) and mounted at /workspace. The workspace
access point is found by root + owner + project tags or created, waited
for before RunTask, recorded store-owned, kept when switching to Hybrid or
Ephemeral, and deleted with the home one by a discarding removal. Hybrid
and Ephemeral task definitions are byte-identical to before (pinned
fingerprints).
The entrypoint links home state for efs as for hybrid, warns when
/workspace is not mounted, and marks a clone in progress in its own
.git/config so a clone a killed task cut short is redone on the next start
while existing repositories are still skipped.
The EFS segment is enabled in HostSection, with a note when /workspace
files remain on an access point after switching away.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
clone_repos now clones into /workspace/.triple-c-clone/<folder> and renames
into place only on success, emptying scratch at each start; nothing at a
final path is ever deleted by the clone code. Replaces the in-repo
triplec.cloning marker, whose failed unset could get a working repository
deleted on the next start. triple-c-wip-push skips the scratch folder.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TRIPLEC_SETUP_SCRIPT runs as claude after the background clone (bash -e,
stdin /dev/null, 30 min timeout) with status/log in /run/triple-c-setup;
the idle watcher counts it as busy and enforces TRIPLEC_MAX_LIFETIME_MINUTES
(best-effort WIP push, marker lifetime, exit 76 = StopKind::LifetimeCap).
WIP pushes go to one triple-c/wip/<run-id> branch per task run, skipping an
unchanged tree. New triple-c-workspace-status prints clone + setup JSON.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two folders sharing a remote no longer force-push over each other's
triple-c/wip/<run-id>; the folder is part of the branch name (sanitised,
check-ref-format validated, hash fallback). A setup chain that dies with
the setup still running records failed so the idle stop is not blocked.
TRIPLEC_MAX_LIFETIME_MINUTES 00 means no cap and 08 means 8. The run record
is written after the agent-token block.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The payload carries TRIPLEC_SETUP_SCRIPT and TRIPLEC_MAX_LIFETIME_MINUTES
(environment hours x 60 when capped); the 64 KiB error names the largest
parts. After Connecting, the start follows triple-c-workspace-status through
the agent (background exec, 2 s, progress on change, 35 min cap, quiet stop
when the tunnel goes). A failed clone/setup is a workspace-warning event and
a warning toast; Overview shows an AWS status card via aws_workspace_status.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ruling R5: the start ends after Connecting (Running, lock released) and a
background follower (cancelled on Stop, Remove, tunnel close, app exit)
reports only a failure it saw or one on a task this start ran. A stale
Starting/Stopping AWS project loads as Error, and reconcile also checks a
Stopped one that still has a task ARN, so a crash never hides a live task.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Copy-at-connect (phase 6, Q2): with secrets_delivery copy_at_connect the
project secret keeps settings, repos and the agent token but drops
TRIPLEC_FILES and every credential pair (SECRET_ENV_KEYS and all custom
variables), marking TRIPLEC_CREDENTIALS=connect. After every connect (start,
attach, background re-attach) the app uploads the SSH/CA files and a
connect.env as root into /dev/shm/triple-c and runs the new
triple-c-apply-connect: CA install as root, SSH merge and git credential
helper as claude, a `connected` marker the first clone waits for, and a
background re-clone of failed repos only. Terminals add the pairs to their
own ExecSpec env. A failed push is a workspace-warning.
The entrypoint's CA, SSH, git-credential, Hybrid-home and clone functions
move to container/entrypoint-lib.sh (sourced first; defines only), so both
scripts run the same code; Docker behaviour is unchanged. HostSection gets
the Credentials choice with its warnings (and the gateway note, R1).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- triple-c-apply-connect: the background re-clone closes the apply lock's
fd, so a later connect never waits on it; --prepare takes the lock.
- R6: under copy-at-connect Hybrid/EFS no longer link ~/.ssh to EFS; the
copied keys vanish with the task (UI says so, and that old keys on EFS
stay until removed).
- R7: a copy-at-connect start that finds credentials still in the secret
force-deletes it and creates it again (retrying ~30 s while AWS finishes),
so no AWSPREVIOUS version keeps them.
- TRIPLE_C_CUSTOM_ENV (unsalted fingerprint) left out of the secret.
- Terminals get only the credential pairs (no file reads; keychain off the
runtime), so a bad SSH/CA file never drops the API keys.
- ConnectBundle values are Zeroizing; UI copy updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Copy-at-connect tries CreateSecret first and reads the current value only
when the secret exists, so a first start never reads a missing secret
(which the owner-tag condition may refuse).
- triple-c-apply-connect: --prepare makes and prints a per-connect nonce
dir; the app uploads into it and applies exactly that dir, so concurrent
connects never mix files. CA files move to a fixed path so an unchanged
CA is not reinstalled.
- connect.env parsing kept a trailing '=' (base64 padding) — it was dropped.
- Pending-deletion errors from GetSecretValue are reported as such; message
and doc-comment fixes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each TCP mapping of an AWS project becomes a loopback-only listener
carried into the task over socat-over-attach (the auth bridge's
PortForward, generalised to separate host and container ports; mappings
skip the bridge's first-byte/web gate). One set per project, opened after
every connect and on a save while connected, closed with the tunnel
(stop, remove, task death, exit). A failing mapping is a workspace-warning;
UDP and reserved ports are refused on save. The Config tab offers the
editor on AWS, TCP only, "Reachable only from this machine (localhost)".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Opens hold a per-project lock and re-read the stored mappings, so
overlapping saves converge on the last one without false 'in use' errors.
Unchanged forwards keep their live connections. The sidecar's 4000 is
refused while it runs (mappings reach workspace loopback). Stored rows
that would be refused no longer block unrelated saves; they are skipped
with a warning at connect. Port-mapping logs get their own prefix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A non-essential 'gateway' container runs the digest-pinned LiteLLM image
next to the workspace (Q7): its secrets are exactly the three
TRIPLEC_GATEWAY_* keys, the config uses os.environ/ references (no key
in it), /bin/sh writes it to a 0600 file and execs LiteLLM on
127.0.0.1:4000. The workspace's ANTHROPIC_BASE_URL/AUTH_TOKEN point at
it with a master key minted per task start; under copy-at-connect that
token and the gateway keys stay in the secret (R1). Needs the
OpenAI-compatible backend and a gateway model (validated on save, start
and in the UI). Requests with the toggle off are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Q5: aws/pricing.rs reads on-demand Fargate rates (x86 and ARM vCPU-hours,
GB-hours, ephemeral storage GB-hours) from the Price List API in us-east-1
(AmazonECS; regionCode, termType=OnDemand, usagetype CONTAINS Fargate),
caches them per region for 24 h under <data_dir>/triple-c/aws-pricing/,
and estimates a config's hourly cost (ephemeral storage billed above
20 GiB; the gateway sidecar adds nothing). aws_cost_estimate(connection_id,
region, config) never errors: denied, offline or slow (8 s) is None,
shown as "Estimate unavailable"; a failed region is not re-asked for
5 minutes and a stale cache still answers.
pricing:GetProducts on * joins workspace-user-tasks.json with its check,
a policy test, the README and a live test (passed). HostSection shows the
estimate while configuring, Overview while running, with the Spot and EFS
notes. Fixture: a recorded us-west-2 GetProducts answer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
a_phase6_workspace_round_trips: one environment (max_workspace_hours 1)
and one project with EFS storage, a setup script, copy-at-connect with a
CA file, a port mapping and the gateway sidecar, two task starts, then
removal with discard. Passed live in us-west-2 (results in aws/README.md
and the spec). The phase-5 setup moves into deploy_for_workspace_user;
aws_host::ensure_started_with takes the sidecar's provider key from the
caller (the live tests have no keychain). scripts/aws-e2e-workspace.sh
runs TRIPLEC_E2E_TEST when set. The guard's task-definition retirement
listed nothing (ECS familyPrefix is a whole family name); it now filters
the full list by prefix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- start_clone_repos waits for the first connect when a setup script exists,
not only when there are repositories, so the script sees the CA, SSH keys
and Git credentials.
- wip_ref appends 8 hex of sha256(raw folder) when sanitising changed the
name, so `a b` and `a_b` never share a WIP branch.
- triple-c-apply-connect moves new CA files to a temporary name and swaps
them in; a failed move keeps the CA installed before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Attaching to a running task no longer needs the provider key or gateway
models; the payload is built lazily on the prepare path only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
begin_follow runs in spawn_follow and the token is passed in, so a quick
Stop→Start can't leave the old follower registered last.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
After a timed-out replace (secret force-deleted, not made again) the next
start retries the create/put for up to 30 s, then says to start again in a
minute; no more advice to restore the secret in the console.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A read in flight blocks the visibilitychange restart, so hiding and showing
the window mid-read no longer starts a second poll chain.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
HostSection gets a monospace "Setup script" textarea (saved on blur, blank
as none) with the 32 KiB UTF-8 limit checked using the backend's message,
and notes on when it runs (as claude, bash -e, after the clone, every start,
30 minutes, output in the Overview card) and what it gets under
copy-at-connect.
Also in Host settings: the environment's advisory lifetime cap (Q8), read
by a new aws_environment_max_hours command, and the copy-at-connect limits
(no CA before the first connect, removed credentials linger in a running
task, no setup re-run after a re-clone), documented in CLAUDE.md too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A save right after the setup script's blur (a toggle) spread the render-time
config and reverted the script; saves now start from the stored config plus
every patch saved since. aws_environment_max_hours is cached per connection,
region and environment for 10 minutes, so re-mounting the section does not
ask AWS again.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
jknapp
merged commit 1659788b70 into main2026-11-16 20:36:37 +00:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Phase 6 of the AWS ECS/Fargate host (spec
docs/superpowers/specs/2026-11-13-aws-ecs-host-design.md, plandocs/superpowers/plans/2026-11-15-aws-phase6.md).What it adds
/workspaceon a second access point of the user's own file system; repos and builds survive a stop.~/.sshstays off EFS; switching an existing project recreates its secret so no old version keeps credentials.claudeafter the clone on every start (30-min limit); progress and failures in the Overview card. Start finishes at "Connecting"; clone/setup are followed in the background and Stop cancels them.max_workspace_hourson the environment (module v2, so existing environments show Update available); enforced inside the workspace (advisory), exit code 76, shown in the project's settings.os.environ/references so no key sits in the task definition.pricing:GetProductstoworkspace-user-tasks.json).Testing
cargo test --locked: 1,506 passed ·vitest: 1,322 passed · build clean · image builds.🤖 Generated with Claude Code